Machine learning for electronically excited states of molecules

Julia Westermayr† and Philipp Marquetand∗,†,‡,¶

†Institute of Theoretical Chemistry, Faculty of Chemistry, University of Vienna, Währinger Str. 17, 1090 Vienna ‡Vienna Research Platform on Accelerating Photoreaction Discovery, University of Vienna, Währinger Str. 17, 1090 Vienna, Austria. ¶Data Science @ Uni Vienna, University of Vienna, Währinger Str. 29, 1090 Vienna, Austria.

E-mail: [email protected]

Abstract

Electronically excited states of molecules are at the heart of photochemistry, photophysics, as well as photobiology and also play a role in material science. Their theoretical descrip- tion requires highly accurate quantum chemi- cal calculations, which are computationally ex- pensive. In this review, we focus on how ma- chine learning is employed not only to speed up such excited-state simulations but also how this branch of artificial intelligence can be used to advance this exciting research field in all its as- pects. Discussed applications of machine learn- ing for excited states include excited-state dy- namics simulations, static calculations of ab- sorption spectra, as well as many others. In order to put these studies into context, we dis- cuss the promises and pitfalls of the involved machine learning techniques. Since the latter are mostly based on quantum chemistry cal- arXiv:2007.05320v1 [physics.chem-ph] 10 Jul 2020 culations, we also provide a short introduction into excited-state electronic structure methods, approaches for nonadiabatic dynamics simula- tions and describe tricks and problems when us- ing them in machine learning for excited states of molecules.

1 Contents 6 Application of ML for Excited States 40 Abstract 1 6.1 Parameters for Quantum Chem- istry ...... 41 1 Introduction 2 6.2 ML of Primary Outputs . . . . . 41 1.1 From Foundations to Applications2 6.3 ML of Secondary Outputs . . . . 41 1.2 Scope and Philosophy of this Re- 6.3.1 ML in the Diabatic Basis . 41 view ...... 3 6.3.2 ML in the Adiabatic Basis 43 6.4 ML of Tertiary Outputs . . . . . 46 2 General Background: From the 6.5 ML-Assisted Analysis ...... 49 Ground State to the Excited States 5 7 Conclusion and Future Perspec- tives 49 3 Quantum Chemical Theory and Methods 8 References 52 3.1 Electronic Structure Theory for Excited States ...... 8 3.1.1 Wave Function Theory 1 Introduction (WFT) ...... 9 3.1.2 Density Functional Theory 13 1.1 From Foundations to Applica- 3.2 Bases ...... 15 tions 3.2.1 Adiabatic (Spin-Diabatic) Basis ...... 16 In recent years, machine learning (ML) has be- 3.2.2 Diabatic Basis ...... 17 come a pioneering field of research and has an 3.2.3 Diagonal Basis ...... 18 increasing influence on our daily lives. Today 3.3 Excited-State Dynamics Simula- it is a component of almost all applications tions ...... 19 we use. For example, when we talk to Siri or 3.3.1 Quantum Nuclear Dy- Alexa, we interact with a voice assistant and namics ...... 20 make use of natural language processing.1,2 ML 3.3.2 Mixed Quantum-Classical is applied for refugee integration,3 for playing Molecular Dynamics . . . 21 board games,4 in medicine,5 for example, for 3.4 Dipole Moments and Spectra . . 22 image recognition6 or for autonomous driving.7 A short historical overview over general ML is 4 Data Sets for Excited States 23 provided in ref 8. 4.1 Choosing the Right Reference Recently, ML has also gained increasing inter- Method for Excited-State Data . 23 est in the field of quantum chemistry.9,10 The 4.2 Phase of the Wave Function . . . 24 power of (big) data-driven science is even seen 4.2.1 Phase Correction of Adi- as the "fourth paradigm of science",11 which abatic Data ...... 25 has the potential to accelerate and enable quan- 4.2.2 ML-Based Internal Phase tum chemical simulations that were considered Correction ...... 27 unfeasible just a few years ago. In general, the 4.3 Training Set Generation . . . . . 28 field of ML in quantum chemistry is progressing 4.3.1 Basic Sampling Tech- faster and faster. In this review, we focus on an niques and Existing emerging part of this field, namely ML for elec- Databases ...... 29 tronically excited states. In doing so, we con- 4.3.2 Active Learning ...... 31 centrate on singlet and triplet states of molecu- lar systems, since almost all existing approaches 5 ML Models 35 of ML for the excited states focus on singlet 5.1 ML Models: Type of Regressor . 35 states and only a few studies consider triplet 5.2 Descriptors and Features . . . . . 38 states.12–15 We note that electron detachment

2 or uptake further leads to doublet and quartet Most ML studies instead focus on predicting states, and even higher spin multiplicities, such the output of a quantum chemical calculation, as quintets, sextets, etc. are common in transi- the so-called "secondary-output".69 Hence they tion metal complexes, where an important task fit a manifold of energetic states of different spin is to identify which multiplicity yields the low- multiplicities, their derivatives and properties est energy and is thus the ground state.15 refs thereof. With respect to different spin states 16–19 give a good overview of such processes. of molecular systems only a few studies exist, The theoretical study of the excited states which predict spins of transition metal com- of molecules is crucial to complement experi- plexes15 or singlet and triplet energies of car- ments and to shed light on many fundamen- benes12 of different composition or focus on the tal processes of life and nature.20 For example, conformational changes within one molecular photosynthesis, human vision, photovoltaics or system13,90,91 for the sake of improving molec- photodamage of biologically relevant molecules ular dynamics (MD) simulations. The energies are a results of light-induced reactions.20–37 of a system in combination with its properties, Experimental techniques like UV/visible spec- i.e., the derivatives, the coupling values between troscopy or photoionization spectroscopy38–45 them, and the permanent and transition dipole lack the ability to directly describe the exact moments,13,14,90–97 can be used for MD simula- electronic mechanisms of photo-induced reac- tions to study the temporal evolution of a sys- tions. The theoretical simulation of the cor- tem in the ground-state 98–135 and in the excited responding experiments can go hand-in-hand states.13,14,90–92,130,136–143,143–147,147–149 with experimental results and can provide the With energies and different properties, ter- missing details of photodamage and -stability tiary outputs can be computed, such as absorp- of molecules.20,27,35,44,46–67 However, the compu- tion, ionization or X-ray spectra,150–153 gaps be- tation of the excited states is highly complex, tween HOMO (highest occupied molecular or- costly, and often necessitates expert knowl- bital) and LUMO (lowest occupied MO) or ver- edge.68 As ML models have only recently been tical excitation energies.154–157 applied in the field of photochemistry, keeping In addition, quantum chemical outputs can track of the approaches is still possible and this also be analyzed or fitted in a direct way, e.g., field is still in its initial stage. reaction kinetics as results of dynamics simu- Due to the multi-faceted photochemistry of lations can be mapped to a set of molecular molecular systems, ML models can target this geometries and can be predicted with ML mod- research field in many different ways, which els.158 Excitation energy transfer properties can are summarized in Figure 1. For example, the be learned,,159,160 and structure-property corre- choice of relevant molecular orbitals for active lations can be explored to design materials with space selections can be assisted with ML.70 The specific properties.16,76,131,152,161–170 fundamentals of quantum chemistry, e.g., to obtain an optimal solution to the Schrödinger 1.2 Scope and Philosophy of this equation or Density Functional Theory, can be central ML applications. For the ground state, Review ML approximations to the molecular wave func- ML for the excited states is developing at a tion71–79 or the density (functional) of a system slower pace than the exploding field of ML for exist.69,79–88 Obtaining a molecular wave func- the electronic ground state.168,171–173 The rea- tion from ML can be seen as the most powerful son is in our opinion mainly a result of the approach in many perspectives, as any property complexity and high expenses of the underlying we wish to know could be derived from it. Un- reference calculations and the associated com- fortunately, such models for the excited states plexity of the corresponding ML models. Sim- are lacking and have yet only been investigated ulation techniques to understand the excited- for a one-dimensional system,89 leaving much state processes are not yet viable for many room for improvement. applications at an acceptable cost and accu-

3 Figure 1: Targets of ML for the excited states of molecules. All areas of excited-state quantum chemistry (QC) calculations can be enhanced with ML, ranging from input to primary outputs that are used in the computation of secondary outputs, which in turn are employed to calculate tertiary outputs. Analysis can be carried out at all stages. This classification is inspired by the one in Ref. 69. racy. Therefore, within this review we also methods with a view to their application in want to highlight the existing problems of quan- time-dependent simulations, namely MD simu- tum chemical approaches that might be solvable lations.27,172 It is worth mentioning, that unlike with ML and put emphasis on identifying chal- for the ground state, where a lot of different lenges and limitations that hamper the appli- methods can provide reliable reference compu- cation of ML for the excited states. The young tations for training, choosing a proper quantum age of this research field leaves much room for chemistry method for the treatment of excited improvement and new methods. states is a challenge on its own. Many methods This review is structured as follows: require expert knowledge, prohibiting their use (1) Throughout this review, we will start further.37,174 In addition, not any method can (non-exhaustively) discussing ground state pro- provide the necessary properties for any type cesses, since they are inherently linked to the of application. Subsequently, we aim to review excited state processes and should also be con- the different flavours of excited-state MD simu- sidered here. We will therefore start by dis- lations with focus on those methods that have cussing the differences between the ground- been enhanced with ML models lately. state potential energy hypersurfaces (PESs) (3) After having provided the basic theoreti- and the excited-state PESs and will also em- cal background, we will discuss how to generate phasize the difference in their properties in a comprehensive, yet full-fledged training set section 2. for the excited states from the quantum chem- (2) Section 3.1 gives an overview of the theo- istry data. We will summarize the existing retical methods that can be used to describe the approaches that are applied to create a com- excited states of molecules. In the forthcoming prehensive training set and put emphasis on the discussion, we will describe different reference bottlenecks of existing methods that can limit

4 also the application of ML. This will provide place in molecules after light excitation, ML the reader with the knowledge about starting models have successfully entered research fields, points for future research questions and clar- which focus on other types of excitations as ify where method development is needed. It well. Those are for example vibrational or ro- further provides the basis for the discussion of tational excitations giving rise to Raman spec- ML models for the excited states of molecular tra or Infrared spectra,43,109,200–204 nuclear mag- systems. netic resonance,205 or magnetism,206,207 which (4) A summary of state-of-the-art ML meth- we will also not consider in this review. ods for photochemistry follows. We will differ- entiate between single-state and multi-state ML models and single-property and multi-property 2 General Background: From ML models.93 As mentioned before, ML models the Ground State to the can tackle a quantum chemical calculation in many different ways, see Figure 1. The different Excited States ML models will be classified in the ways they enhance quantum chemical simulations. Most approaches aim at providing an ML-based force field for the excited states, so most focus will be put on this topic. At last, the prospects of ML models to revolutionize this field of research and future avenues for ML will be highlighted.

Noteworthy, we focus on the excited states of molecules, as the excited electronic states in the condensed phase are challenging to fit and are thus often not explicitly considered in conven- tional approaches.175–180 In solid state physics Figure 2: Excited-state processes that can take for example, the electronic states are usually place after excitation of a molecule by light. treated as continua. The density of states at Absorption of light can make the molecule enter the Fermi level,181 band gaps,182–184 and elec- a higher electronic singlet state. Intersystem tronic friction tensors123,185,186 have been de- crossing to a triplet state or internal conver- scribed with ML models up to date and es- sion to another state of same spin-multiplicity pecially the electronic friction tensor is use- can take place. Radiative emission, i.e., fluo- ful to study the indirect effects of electronic rescence and phosphorescence, are possible re- excitations in materials.187–192 Electron trans- actions from an excited singlet and triplet state, fer processes as a result of electron-hole-pair respectively. excitations can be further investigated along with multi-quantum vibrational transitions by Figure 2 gives an overview of the excited state discretizing the continuum of electronic states processes that will be discussed within this re- and fitting them (often manually) to reproduce view. It shows a schematic one-dimensional experimental or quantum chemical data in a representation of the potential energy curves 178,193–198 model Hamiltonian. Yet, to the best of for the ground and excited states as a func- our knowledge, the excited electronic states in tion of molecular coordinates. Figure 2 il- the condensed phase have not been fitted with lustrates that the ground state potential en- ML. A recent review on reactive and inelastic ergy curve, given by a dark-blue solid line, is scattering processes and the use of ML for quan- mostly a smooth function of the reaction coor- tum dynamics reactions in the gas phase and at dinate and gives information about several lo- a gas-phase interface can be found in ref 199. cal minima. In the ground state, many meth- Besides the electronic excitations that take ods exist to describe the physico-chemical prop-

5 erties of molecules and materials reasonably can overcome the limitations of existing force well, ranging from small systems up to proteins, fields.133,171,223–227 DNA or nanoparticles. For small system sizes, Regarding the excited states, processes be- highly accurate ab-initio methods can be ap- come much more complex and the computation plied, while more crude approximations have of excited state PESs is far more difficult than to be used for larger systems. The unfavor- the computation of the ground state PESs. As able scaling of many quantum chemical meth- can be seen in Figure 2, a lot of different classes ods with the size of system under investigation of excited states, e.g. singlet states as shown requires this compromise between accuracy and by continuous blueish lines or triplet states as system size. Crude approximations for systems shown by dashed reddish lines, have to be ac- that are larger than several 100s of atoms be- counted for, which are characterized by several come inevitable.20,24,37 transition states, local minima, and crossing The chemistry we are interested in, however, points. This complexity makes a separate treat- is not static, but rather depends to a large ment of each electronically excited state inac- extent on the changes that matter undergoes. curate and leads to further challenges that pro- In this regard, it is more intuitive to study hibit the straight-forward and large-scale use of the temporal evolution of a system. Much ef- many existing quantum chemical methods and fort has been devoted to develop methods to consequently also existing ML models for the study the temporal evolution of matter in the ground state. ground state potential. As an example, physi- Additionally, computations of the excited cal functions can be obtained with conventional states suffer from being generally less efficient. force fields, such as AMBER,208 CHARMM209 To name only one central problem: The larger or GROMOS.210,211 The first ones already date the system becomes, the closer the electronic back to the 1940s-1950s. Such force fields states lie in energy, and the more excited-state enable the study of large and complex sys- processes can usually take place. The neces- tems, protein dynamics or binding-free ener- sary consideration of an increasing number of gies on time scales up to a couple of nanosec- excited states increases the already substantial onds.175,212–220 However, their applicability is computational expenses even more and restricts restricted by the limited accuracy and inabil- the use of accurate methods to systems contain- ity to describe bond formation and breaking. ing only a few dozens of atoms in a reasonable Novel approaches, such as reactive force fields amount of time with current computers. This exist, but have not yet entered the mainstream increasing complexity makes not only the refer- and still face the problem of generally low ac- ence computations, but also the application of curacy.221 ML models for the excited states more compli- The accuracy of ab-initio methods can be cated than for the ground state. At the same combined with the efficiency of conventional time, the application of ML models for the ex- force fields with ML models. The latter have cited states might also be more promising, be- shown to advance simulations in the ground cause higher speed-ups can be achieved. state considerably and allow for the fitting of al- For the excited states, methods similar to most any input-output relation.98–132,135,172,222 force fields, like the linear vibronic coupling Accurate and reactive PESs of molecules in the (LVC) approach,228,229 are usually limited to ground state can be obtained with a compre- small regions of conformational space and re- hensive reference data set, which contains the stricted to a single molecule. General force energies, forces and ground-state properties of a fields that are valid for different molecules in system under investigation. Proper training of the excited states do not exist. Also the ML an ML model then guarantees that the accuracy analogue, so-called transferable ML models, to of the reference method is retained, while infer- fit the excited state PESs of molecules through- ences can be made much faster. In this way, out chemical compound space are unavailable they allow for a description of reactions and up to date. Nevertheless, it is out of ques-

6 234 tion that an ML model, which is capable of de- ence, ∆Eij: scribing the photochemistry of several different molecular systems, e.g., different amino acids or osc 2 2 fij = ∆Eij | µij | . (1) DNA bases of different sizes, is highly desirable. 3 A lot remains to be done in order to achieve If the transition dipole moment between two this goal and yet, to the best of our knowledge, states is zero, no transition is allowed. The no more than a maximum of about 20 atoms reasons can be that a change of the electronic and 3 electronic states with a distinct multiplic- spin would be required, and the transition is ity have been fitted accurately with ML mod- thus spin forbidden. Another reason can be 13,14,90–92,94,130,136–143,143,144,144–147,147–149 els. the molecular symmetry, leading to symme- Whether or not the excited states of a molec- try forbidden transitions. The latter are com- ular system become populated depends on the mon in molecules that carry an inversion centre ability of a molecule to absorb energy in the and transitions that conserve parity are forbid- form of light, or more generally, electromag- den.235 An energetic state is called dark, if the netic radiation of a given wavelength. Usu- transition dipole moment is very small or zero. ally, the so-called resonance condition has to be In contrast, a state is called bright, if the transi- fulfilled, i.e., the energy gap between two elec- tion dipole moment is large. Most often, studies tronic states has to be equivalent to the photon that target the photochemistry of molecules fo- energy of the incident light. Note however that cus on excitation to the lowest brightest singlet also multi-photon processes can occur, where state, i.e., the state that absorbs most of the several photons have to be absorbed at once to incident energy. bridge the energy difference between two elec- After an excitation process, the molecule is 230–232 tronic states Further, the absorption of considered to move on the excited-state PESs light does not only provide access to one, but and is expected to undergo further conversions. most often to a manifold of energetically close- The excess of energy a molecule carries – as a lying states. The number of states that can be result of the initial absorption of energy – is excited is related to the range of photon en- most often converted into heat, light, such as ergies that is contained in the electromagnetic fluorescence or phosphorescence, or into chem- radiation. This energy range is inversely pro- ical energy. If the molecule returns to its orig- portional to the duration of the electric field, inal state, then the molecule is photostable. e.g., of a laser pulse, due to the Fourier rela- Otherwise, either photodamage, such as de- 233 tion of energy and time. However, the en- composition, or useful photochemical reactions ergy range of the photons and the energy dif- including bond breaking/formation occur. In ference between the electronic states are not all cases, heat or light can be emitted, which the only factors influencing the absorption of can also be harnessed in light-emission applica- light, which gives rise to questions like: Is the tions.27,236–238 With respect to photo-stability, molecule able to absorb light of a considered ultrafast transitions, in the range of femto- to wavelength? Which of the excited states is pop- picoseconds (10−15–10−12 seconds) take place ulated with the highest probability? and lead the molecule back to the ground state. An answer to these questions can be obtained This means, the electronic energy is converted from an analysis of the oscillator strength. In into vibrations of the molecule and the molecule order to make an electronic transition possible, is termed hot. This heat is usually dissipated an oscillating dipole must be induced as a re- into the environment, a procedure that is of- sult of the interaction of the molecule with light. ten neglected in excited-state simulations due osc The oscillator strength, fij , between two elec- to the cost of describing surrounding molecules. tronic states, i and j, is proportional in atomic Radiationless transitions from one electronic units (a.u.) to the respective transition dipole state to another take place in so-called criti- moment, µij, and the respective energy differ- cal regions of the PESs. As the name already

7 suggests, critical regions are crucial for the dy- states. These challenges also point at issues namics of a molecule, but are also challeng- that are problematic for ML. These explana- ing to model accurately. A transition from tions will provide the groundwork to evaluate one state to another that conserves the spin- different quantum chemical methods for their multiplicity is called internal conversion. Fur- use to generate a training set for ML and to thermore, states of different spin-multiplicities use it for different types of applications, such may be accessible via intersystem crossing. The as excited-state MD simulations. Naturally, we critical points, where transitions are most likely can only provide a general idea of this field and to occur, are called conical intersections and are refer the interested reader to pertinent text- illustrated in Figure 2. At these crossing points, books and reviews, such as Refs.26,29,36,37,240–248 PESs computed with quantum chemistry can In order to follow a consistent notation within show discontinuities. These discontinuities can this review, we try to explain all basic con- occur also in other excited-state properties and cepts with notations that are frequently used pose an additional challenge for an ML model in literature. Currently, a zoo of different no- when fitting excited-state quantities. tations for the same property can be found. In addition to the aforementioned complica- For example, the NACs, or derivative couplings, tions of treating a manifold of excited states, are sometimes referred to as so-called inter- also the probability of a radiationless transi- state couplings, i.e. couplings between two tion between them has to be computed some- states multiplied with the corresponding energy how. This probability is usually determined gap between those two states,142 while in other by couplings between two approaching PESs. works interstate couplings refer to off-diagonal Between states of the same spin multiplicity, elements of the Hamiltonian in another basis, nonadiabatic couplings (NACs) arise, and spin- where the potential energies are no eigenval- orbit couplings (SOCs) give rise to the transi- ues of the electronic Schrödinger equation. We tion probability between states of different spin want to avoid a confusion of the different no- multiplicities. These couplings are intimately tations and thus provide a consistent definition linked to the excited-state PESs and therefore below. For the excited states, a number of dif- should also be considered with ML. However, ferent electronic states is required. Throughout only a handful of publications describe cou- this review, we adopt the following labelling plings with ML,13,90,92–94,138,143,144,147,239 which convention for different electronic states: The highlights the difficulty of providing the neces- lower case Latin letters, i, j, etc. will be used sary reference data as well as the challenges of to denote different electronic states. The ab- accurately fitting them. New methods are con- breviations NS, NM , and NA will indicate the stantly needed to further enhance this exciting number of states, molecules and atoms, respec- research field. tively. The foundation for the following sections is a separation of electronic and nuclear de- 3 Quantum Chemical The- grees of freedom, which is based on the work ory and Methods of Born and Oppenheimer.249 However, the famous Born-Oppenheimer approximation is In this section, we present some key aspects of later on (partly) lifted and the coupling be- quantum theory for excited states because (i) tween electrons and nuclei is taken into account the outcome of the corresponding calculations in nonadiabatic dynamics simulations. serve as training data for ML and (ii) to clar- ify the employed nomenclature. We put spe- 3.1 Electronic Structure Theory cial emphasis on describing the differences of for Excited States excited-state computations to computations in the ground state and the challenges that arise The main goal when carrying out an electronic due to the treatment of a manifold of excited structure calculation is usually to compute the

8 potential energy and other physico-chemical accurately has not yet been found. Moreover, properties of a compound. We distinguish be- there is no systematic way to improve a den- tween two overarching theories to achieve this sity functional. The results obtained with DFT goal: Wave Function Theory (WFT) and Den- therefore critically depend on the choice of the sity Functional Theory (DFT) – as outlined, functional.253,255 e.g., by Kohn in his Nobel lecture.250 In the following sections, we will describe The basis of WFT, as for any electronic struc- both theories in the light of excited states of ture calculation, is the electronic Schrödinger molecules. We will start to cover ab-initio equation251,252 with the electronic Hamilton op- methods, which means that they are derived ˆ erator, Hel, and the N-electron wave function from first principles without parametrization. Ψi(R, r) of electronic state i, which is depen- dent on the electronic coordinates r and para- 3.1.1 Wave Function Theory (WFT) metrically dependent on the nuclear coordi- nates, R: The basis of all discussed ab-initio methods is the Hartree-Fock method. The N-electron ˆ Hel(R, r) | Ψi(R, r)i = Ei | Ψi(R, r)i. (2) wave function is represented by a single Slater Determinant, φ0, which makes N coupled one- From the wave function, the eigenvector of this electron problems out of the N-body problem. eigenvalue equation, any property of the sys- This Slater determinant is the anti-symmetric tem under investigation can be derived. How product of one-electron wave functions, the spin to solve the electronic Schrödinger equation ex- orbitals, which can be atomic, molecular or actly to obtain the potential energy of an elec- crystal orbitals, depending on the system. In tronic state i, Ei, is known in theory. However, the case of molecular (or also crystal) orbitals, from a practical point of view, the computa- they are usually expanded as a linear combina- tion is infeasible for molecules that are more tion of atomic orbitals, where the expansion co- + complex than for example H2, He2 , and similar efficients are optimized during the calculation. systems.253 In order to make the computation In order to do so efficiently, the atomic orbitals of larger and more complex systems viable, ap- are themselves expanded with the help of a ba- proximated wave functions are introduced. sis set. The N-electron wave function is there- In contrast to WFT, DFT reformulates the fore obtained as a double expansion. Two ap- energy of a system in terms of the ground state proximations are applied, which is the use of a electron density rather than the N-electron finite basis set to represent the atomic orbitals wave function and the energy is expressed as a and in turn also the molecular orbitals on the functional thereof. The advantage of DFT over one hand and the use of a single Slater Deter- WFT is a rather high accuracy for a rather low minant on the other hand. This usually gives computational cost. If DFT is applied prop- a poor description of a system under investiga- erly, it is considered as one of the most efficient tion, due to a lack of electronic correlation. ways to obtain reliable and reasonably accurate Electronic correlation describes how much the results of molecules up to 100s of atoms. In motion of an electron is influenced by all other solid state physics, DFT is even the workhorse electrons. Since the Hartree-Fock method can of most studies aiming to describe ground state be seen as a mean-field theory, where an elec- properties.254 However, the problem is that the tron ”feels” only the average of the other elec- equations to be solved are unknown. The miss- trons, correlation is quantified by the correla- ing piece is the exact exchange-correlation func- tion energy, which is the difference between the tional of a system. Up to date, researchers have Hartree-Fock energy and the exact energy of a come up with many different approximations to system. this functional that can be used to treat spe- Unsurprisingly, all further discussed quan- cific problems, but a universal functional ca- tum chemical methods aim at improving the pable of describing different problems equally Hartree-Fock method. They can be seen as dif-

9 ferent flavors of the same solution to the prob- sis set. The use of all possible configurations is lem: They all include more determinants in called Full-CI and represents the case, when all one way or another. Accordingly, the wave electrons are arranged in all possible ways. This function is expanded as a linear combination approach is infeasible for almost all molecular of determinants, where a determinant consists systems, more complex than e.g. He, and trun- of molecular orbitals, which are expanded in cated methods are needed. Those are for exam- atomic orbitals. This ansatz contains two ple, CIS (CI Singles) or CISD (CIS and Dou- types of coefficients that can be optimized, bles), where only single excitations or addition- the ones for the determinants and the ones ally double excitations are accounted for, re- yielding the molecular orbitals. If the latter spectively. Figure 3 gives a schematic overview are kept the same for different determinants, of the improvements of CI that one can apply. we speak of a single-reference wave function. A huge advantage of these methods is, that how If both types of coefficients are adapted, we to obtain the exact solution is known, and that speak of a multi-reference wave function. Sim- they are systematically improvable. However, ilarly, the electron correlation is also divided truncated CI does not scale correctly with the into two parts, termed dynamic correlation and system size and is therefore not size-extensive static correlation. Single-reference methods im- and also not size-consistent (i.e., the energy of prove on the dynamic correlation, while a multi- two fragments A and B at large distance com- reference wave function allows for static corre- puted together, E(A + B), is not equal to the lation. However, the separation is not so strict, sum of the energies of the fragments from sep- as can be seen by the following fact: Both the arate calculations, 6= E(A) + E(B)).258 aforementioned single-reference variant and the multi-reference variant become equivalent when including an infinite number of terms and de- liver the exact solution to the Schrödinger equa- tion if also an infinite basis set is used.

Configuration Interaction In the case of single-reference methods, the orbitals obtained from the reference calculation (usually Hartree- Figure 3: Different arrangements of electrons in Fock) are kept fixed. Since usually more or- molecular orbitals giving rise to the configura- bitals than the number of electrons in the sys- tion interaction (CI) method. Inclusion of ex- tem are calculated, the possibility of construct- cited configurations in addition to the ground- ing different Slater determinants from these or- state, reference determinant, φ0, allows to go bitals exist, which can be used for expanding beyond the Hartree-Fock method. Electrons the actual wave function:256,257 are excited into higher electronic orbitals and X Slater Determinants are indicated using the let- | Ψii = cI | φI i (3) ters S, D, T, and Q, which refer to single, dou- I ble, triplet, and quadruple excitations. Each Slater Determinant is weighted by a co- The CI scheme can be employed to improve efficient, cI . These coefficients can be obtained variationally by minimizing the total energy un- the ground-state wave function by mixing the der the constraint of fixed orbitals, ending up Hartree-Fock determinant and determinants of in the Configuration Interaction (CI) methods. different electron configurations. In the same way, also wave functions of excited states can Ψ0 is the reference, Hartree-Fock, wave func- tion. In principle, the exact solution can be be computed. Then, the coefficients CI are op- obtained by considering all possible Slater De- timized for higher eigenvalues of the electronic terminants in combination with a complete ba- Hamiltonian instead of the first one. Beginners in the field then often get confused by terms

10 like single excitation in comparison to first ex- operator, Tˆ:266 cited state. A single excitation determinant Tˆ (see Fig. 3) can be part of the wave function | φCC i = e | Ψ0i = for the first excited state but can also be a part 1 1 (4) (Tˆ = 1 + Tˆ + Tˆ2 + Tˆ3 + ...) | Ψ i. of the ground-state wave function. 2! 3! 0 Similarly to CI, this operator can be truncated. Electron Propagator Methods Another If Tˆ = Tˆ + Tˆ , single and double excitations class of methods that we shortly want to men- 1 2 are accounted for. tion here are electron propagator methods, that Excited states can be computed in a single- are based on one electron Green’s function reference approach by equation-of-motion-CC and are another variant of perturbation theory (EOM-CC), where the excited-state wave func- schemes. One popular method that is based tion is written as an excitation operator times on Green’s function one electron propagator ap- the ground-state wave function. For further de- proach is the algebraic diagrammatic construc- tails, see, e.g., the reviews 267,268. tion scheme to second order perturbation the- ory (ADC(2)).259 ADC(2) is a single-reference method and can be used to efficiently compute CASSCF The problem of missing static cor- excited states of molecules. It offers a good relation in the Hartree-Fock approach is tackled by a multi-reference ansatz for the wave func- compromise between computational efficiency 255 and accuracy, while being systematically im- tion. This treatment is important for many provable (higher order variants like ADC(2)-x excited-state problems, but also some transition or ADC(3) exist). The time evolution of a sys- metal complexes in their ground state, transi- tems polarizability is obtained by applying the tion states or homolytic bond-breaking with the dissociation of the N2 molecule being a notori- polarization propagation, which contains infor- 269,270 mation on a system’s excited states.256,260–263 ously difficult example. The ground-state energy of ADC(2) is based The multi-configurational self-consistent field on Møller-Plesset perturbation theory of sec- (MCSCF) method can be seen as the multi- 264,265 reference counterpart to the Hartree-Fock ond order, MP2, where the latter can for- 271 mally be shown to include double excitations for method. One of the most popular variants of MCSCF methods is the Complete Active Space the improvement of Hartree-Fock, see Ref. 256. 272,273 The dependence of ADC(2) on MP2 gives rise SCF (CASSCF), where important atomic to instabilities in regions, where excited states orbitals and electrons are selected giving rise to come close to the ground state or homolytic an active space. An example is shown in Fig- dissociation takes place. The excited states ure 4. According to this scheme, the orbitals of bound molecules are described with reason- are split into an inactive, doubly occupied part, able accuracy. Compared to multi-reference CI an active part and an inactive, empty part. methods (see below), the black box behaviour Within the active space a FCI computation is of ADC(2) is a clear advantage.259 carried out. The active space has to be chosen manually by selecting a number of active elec- trons and active orbitals. CASSCF is no black Coupled Cluster The gold standard of ab- box method and a meaningful active space se- initio methods for the ground state is the fam- lection is the full responsibility of the user. As ily of Coupled Cluster (CC) methods. CC is an advantage, CASSCF can describe static cor- often referred to as the size-extensive and size- relation well, which is necessary in systems with consistent version of CI. The different electronic nearly degenerated configurations with respect configurations accounting for single or double to the reference Slater determinant. For com- excitations (such as in CIS and CISD for ex- pleteness, state-averaging (i.e. SA-CASSCF) ample) are obtained by applying an excitation is most often applied, where states belonging to the same symmetry are averaged. Another

11 problems exist, like the n-electron valence state perturbation theory (NEVPT2).279–281

MRCC In addition to multi-reference meth- ods based on CI, multi-reference variants of CC approaches exist. A relatively efficient imple- mentation is for example the Mk-MRCC ap- proach of Mukherjee and co-workers282 or the Brillouin-Wigner approach,283 which is however Figure 4: Electrons and orbitals of an arbitrary not size extensive. Noticeably, the development system to exemplify the active space needed for of multi-reference CC approaches is a rather many multi-reference methods. (a) The high- young research field compared to other excited- est, not considered, molecular orbitals are in- state methods and the computation of proper- active and always empty. (c) The lowest, not ties and forces is not well explored. Many stud- considered, molecular orbitals are always dou- ies therefore focus on the simulation of energies bly occupied. (b) The active space is shown of low-lying states with MRCC methods. Ad- with two active electrons in two active orbitals. ditionally, such methods suffer from algebraic The occupancy of the orbitals is between zero complexity and numerical instabilities. Inter- and two. ested readers that seek for a more extensive summary of existing MRCC methods are re- variant of MCSCF methods is restricted active ferred to Refs. 29,284,285. space SCF (RASSCF), which is very similar to CASSCF, but within RASSCF the active space Challenges The probably biggest drawback is restricted and no FCI computation is carried of the aforementioned multi-reference meth- out.256 ods is that their protocols are very demand- ing. Finding a proper active space is a te- MR-CI Even higher accuracy can be dious task that often requires expert knowledge. obtained with multi-reference CI meth- Too small active spaces can lead to inaccurate ods,29,274,275 such as MR-CISD, that addition- energies and problems with so-called intruder ally add single and double excitations out of states are common. Those are electronic states, the active space and are therefore based on that are high in energy at a reference molecu- CASSCF wave functions. With this approach lar geometry, but become very low in energy electronic correlation, i.e. static and dynamic at another molecular geometry, that is visited correlation, can be treated. along a reaction coordinate. The active space then changes along this path. This behavior CASPT2 Alternatively, complete-active- can result in inconsistent potential energies. In space perturbation theory of second order, case of CASPT2, the configurations of intruder CASPT2,276–278 can correct electronic corre- states can lead to large contributions in the lation effects via treating multi-reference prob- second-order energy, making the assumption of lems with perturbation theory. This variant small perturbations invalid. Especially for de- of multi-reference perturbation theory methods scribing molecular systems with many energet- uses the CASSCF wave function as the ze- ically close-lying states and for the generation roth order wave function. CASPT2 can be ap- of a training set for ML, such inconsistencies plied to each state separately (single-state (SS)- are problematic. Figure 5 shows an example CASPT2) or correlated states can be mixed at of potential energy curves of 3 singlet states second order resulting in a multi-state pertur- and 4 triplet states of tyrosine computed with bation treatment (MS-CASPT2).276–278 Other (a) CASSCF(12,11) and (b) CASPT2(12,11), perturbation approaches for multi-reference where 12 refers to the number of active elec- trons and 11 to the number of active orbitals.

12 We used OpenMolcas286 to compute an unre- needs to be large enough to treat both, the va- laxed scan along the reaction coordinate, which lence and Rydberg molecular orbitals. Addi- is a stretching of the O-H bond located at the tionally, the one electron basis set should be phenyl-ring of tyrosine. flexible enough to describe both types of or- bitals. This increases the computational costs additionally. More details on the inclusion of Rydberg states in simulations can be found in refs 289–292. A promising tool to eliminate the complex choice of active orbitals is autoCAS.293–295 It provides a measure of the entanglement of molecular orbitals that is based on the den- sity matrix renormalization group (DMRG). A DMRG-SCF calculation is similar to a CASSCF calculation, but instead of a FCI solution of the active space, an approximated solution with DMRG is obtained to avoid the exponential scaling of the computational costs with the Figure 5: Potential energy curves of the three number of active orbitals.296–301 As an alterna- lowest singlet (S0-S2) and the four lowest triplet tive, ML can be used to determine an active state (T1-T4) of the amino acid tyrosine along space.70 the O-H bond length of the hydroxy group located at the phenyl ring (Ph-OH) com- 3.1.2 Density Functional Theory puted with CASSCF(12,11)/ano-rcc-pVDZ and CASPT2(12,11)/ano-rcc-pVDZ.287 A complementary view on how to obtain the energy of a system is provided by DFT. DFT Intruder states are no exception. Actu- dates back to 1964, when it was formulated by ally, they are quite common in small to Hohenberg and Kohn302 entirely in terms of the medium sized organic molecules. A large electron density, η(~r). A one-to-one correspon- enough reference space can mitigate this prob- dence between this density and an external po- lem, but makes computations almost infeasi- tential, v(~r), exists and the potential acts on the ble. The computational costs increase expo- electron density. The energy can be formulated nentially with the number of active orbitals. in terms of a universal functional, F [η(~r)], of In many cases, the improved accuracy due to the electron density, which is independent of the a larger active space cannot justify the consid- external potential. In this way, the energy of a erably higher expenses. At its best and with system’s ground state can be computed with massively parallel simulations, an active space the following equation: of about 20 electrons in 20 orbitals can be Z treated,288 which is impracticable for many ap- E[η(~r)] = v(~r)η(~r)d~r + F [η(~r)] (5) plications, such as dynamics simulations. For medium-sized molecules, the active space that The most widely used implementations of DFT would be required for a given simulation might rely on the Kohn-Sham approach.303 In fact, even be way to large to be feasible for calcula- Kohn-Sham DFT is so successful that it is often tions in a static picture. simply referred to as DFT. In this approach, an Worth mentioning at this point are also Ry- auxiliary wave function in the form of a Slater dberg states, that often need to be considered determinant is employed. Since a single Slater in small to medium sized molecules. Rydberg determinant is the exact solution for a system states can be strongly interlaced with valence of noninteracting electrons, this DFT approach excited states. In such cases, the active space can be seen as describing a system of nonin-

13 teracting electrons that are forced to behave as time-dependent electron density in this poten- if they were interacting. The latter effect can tial. A system can therefore be completely de- be achieved only by an unknown modification scribed by its time-dependent density. Also in of the Hamiltonian or rather of the aforemen- the time-dependent case, the variational princi- tioned functional. In other words, a Slater de- ple for the density is proposed. terminant as wave function ansatz is exact but The most widely used approach of TDDFT is the Hamiltonian can only be approximated, in linear response TDDFT (LR-TDDFT). Again, contrast to Hartree-Fock, where the true elec- often TDDFT is used synonymously with LR- tronic Hamiltonian is used but the Slater deter- TDDFT due to its extensive use. Within this minant is only an approximate wave function. theory and the KS approximation, no time de- The functional F [η(~r)] can be separated into pendent density is necessary to compute ex- Coulombic interactions and a non-Coulombic citation energies and excited state properties. part. The latter can further be divided into two Linear response theory can be directly applied terms: the kinetic energy of the noninteract- to the ground state density.306,307 Casida’s for- ing electrons and the exchange-correlation part, mulation of this theory is the most popular which describes the interaction of electrons and one and gives rise to random-phase approxi- thus also corrects the kinetic energy by the dif- mation pseudo-eigenvalue equations, which are ference of the real kinetic energy and the kinetic also known as the Casida equations. Within energy of the fictitious system of noninteracting the adiabatic approximation, they are imple- electrons. The exchange-correlation functional mented efficiently in many existing electronic is the part of DFT that is unknown and finding structure programs. The Tamm-Dancoff ap- the exchange-correlation functional remains the proximation308,309 further simplifies the equa- holy grail of DFT. tions to an eigenvalue problem, resulting in In principle, if the exact functional was the counterpart to CIS.310 Especially in cases, known, the exact ground-state energy of a sys- when the time evolution of a system is studied, tem could be computed. Unfortunately, it is the Tamm-Dancoff approximation is beneficial, not known and the success of a DFT calculation since it leads to more stable computations close critically depends on the approximation that is to critical regions of the PESs.253,304 used to the unknown exchange-correlation func- The advantage of LR-TDDFT is its computa- tional. For completeness, KS-DFT is often used tional efficiency. The reasonable accuracy if a for closed-shell systems. In case of open-shell proper functional is chosen makes this approach systems, two spin densities are distinguished, often the method of choice to study the photo- resulting in spin-polarized KS theory.304 chemistry of medium-sized to large and com- As explained above, the electron density is plex systems, which are not feasible to treat computed from a single reference Kohn-Sham with costly multi-reference WFT based meth- wave function, i.e., the one of noninteracting ods.29,311,312 Shortcomings of LR-TDDFT are electrons with the density of the real system. the incorrect dimensionality of conical intersec- This single-reference wave function makes DFT tions, which are, however, one of the most im- a single-reference method. In fact, most failures portant regions during nonadiabatic MD sim- of DFT are a consequence of an improper de- ulations.313–315 The incorrect dimensionality of scription of static correlation.255 In order to de- conical intersections with standard TDDFT im- scribe excited states, the time-dependent (TD) plementations leads to a qualitatively incor- version of DFT, namely TDDFT, can be used. rect description of such critical regions. The The foundation of this theory was laid in the missing couplings can be corrected for example 1980s with the Runge-Gross theorems,305 which with the CI-corrected Tamm-Dancoff approxi- can be regarded as analogies to the Hohenberg- mation316 or the hole-hole Tamm-Dancoff ap- Kohn theorems. They are based on the assump- proximation,317 which can recover the missing tion that a one-to-one correspondence exists couplings and provide correct dimensionality at also between a time-dependent potential and a conical intersections.

14 In addition, one should be aware that by def- to a particular problem, but that many possi- inition, double excitations cannot be accounted ble ways can be considered which lead to an for with LR-TDDFT. The computation of dou- equivalent description of a particular problem. ble excitations can be achieved by using a fre- Considering the excited states of molecules, quency dependent exchange kernel, which is it should be mentioned that it is of utmost known as dressed TDDFT.318,319 Alternatively, importance to think carefully about the photo- spin-flip TDDFT320,321 can be used, where a chemical processes that may occur in order to triplet state is taken as a reference state and find the most appropriate method for most of single excitations are treated with a flip in the the assumed reactions. It often happens that electron’s spin. However, spin-contamination is within the same molecular system, one method quite common within these methods. In gen- can describe a certain photochemical reaction eral, the description of double excitations from quite well, while another reaction can be de- a multi-reference state would be more favorable, scribed better with another method. However, although spin-flip TDDFT is often considered the mixing of methods is not practicable for to be a multi-reference method. In order to standard applications. Recently, studies on ML compute specific orbital occupations and conse- models have emerged that combine the different quently excitations and charge-transfer states, strengths of several methods, e.g. ∆-learning an alternative approximation exists, which is techniques327,328 or transfer learning.329 These known as the ∆-SCF approach. In this the- methods could be well-suited solutions for many ory, the electrons are forced into specific KS future applications to overcome the current lim- orbitals. The SCF is applied to converge the itations of existing quantum mechanical meth- energy with respect to this configuration.322–324 ods for the excited states. Even more than Other multi-reference variants of TDDFT exist for ground state properties, the quality of the too. However, their description is beyond the excited states depends critically on the ability scope of this review and we refer the reader to of a method to describe the different possible a review covering this topic in much more de- reactions - as a consequence of the larger acces- tail.29 sible configuration space of a molecular system. Last but not least, we shortly want to discuss Even for medium-sized systems it should be the most critical part of a DFT calculation, clear that a suitable method may already be which is the proper choice of the exchange- computationally impracticable and a balance correlation functional. In case of excited states, between accuracy and computational effort has the treatment of valence excitations, Rydberg to be found. states and long-range charge transfer excita- tions on the same footing is highly problematic. 3.2 Bases While hybrid (meta-) generalized gradient ap- proximation (GGA) or range-separated hybrid The potentials computed with the aforemen- functionals325 are for example well suited for tioned methods for different nuclear geometries vertical excitations and the latter also for Ryd- can be represented in different bases, which are berg states, global hybrid meta GGA or range- connected by unitary transformations. An ex- separated hybrid GGA functionals are better ample of five states in different bases are given to describe charge transfer.253,326 Most often, in Figure 6. Note that often a system in a cer- functionals are accurate for one specific prob- tain basis is also referred to as being in a certain lem, but they fail to describe others. Although picture or representation; here we will not use much effort has been devoted to develop func- the term representation in order to not confuse tionals, finding a universal functional for DFT the reader with molecular representations used is still far from being achieved.29,175,253,304 in ML. As it is visible in the figure, we focus on three types of bases: (a) the diabatic basis, In summary, it should be stressed that, in (b) the adiabatic (spin-diabatic) basis, i.e., the general, there is not only one single solution direct output of standard electronic structure

15 programs, (c) the diagonalized version of (a) a=not, dia=through, batos=passable) and, in- and (b), i.e., the spin-adiabatic basis. Through- deed, the potentials never cross when consider- out literature, different names are given to these ing one multiplicity. This situation is schemati- bases, which are summarized in Table 1. They cally illustrated in Figure 6(b) for singlet Si and stem from a partition of the total wave function singlet Sj. into a sum of electronic and nuclear contribu- Within one multiplicity, 3NA-dimensional tions, which can be written for all bases as: adiabatic PESs are obtained that are strictly ordered by energy. Hence, the states are usu- X basis basis Ψ(r, R, t) = ψi (r, R)χi (R, t). (6) ally denominated with the first letter of the i multiplicity and a number as subscript, e.g., S ,S , etc. For states of the same multiplicity, In a similar way as the number 20 can be fac- 0 1 critical points and seams exist. These regions tored into 4·5 or 4.5·4.4¯, the total wave function of the PESs are referred to as conical intersec- can be expanded in the different bases. Here, tion (seams), in which the corresponding states ψbasis(r, R) corresponds to the eigenfunctions of i become degenerate. Such features make adia- the electronic Hamiltonian only for one of the batic PESs non-smooth functions of the atomic bases (namely the one from column B of ta- coordinates, which make them difficult to pre- ble 1). Associated with these functions are the dict with the intrinsically smooth regressors of corresponding potentials, depicted for a model ML. At a conical intersection, the approaching system in Fig. 6. Note that a different approach potential energy curves form a cone and the is taken in the exact factorization method,330 NACs, denoted as C , between them show where the total wave function is expanded only NACij singularities as a result of the inverse propor- in a single product, i.e., without the sum in tionality to the vanishing energy gap:274,335 eq. 6, giving rise to only one (time-dependent) potential. ∂ CNACij ≈ hΨi | ∂R Ψji = 1 ∂Hel (7) hΨi | | Ψji for i 6= j, Table 1: Commonly used names of bases for the Ei−Ej ∂R excited-state potential energy surfaces based on Second order derivatives are neglected here, as refs 228,242,331–334. The labels a, b, and c are is done in many quantum chemistry programs consistent with Fig. 6 that compute NAC vectors. The blue dashed a b c curve in panel (b) of Fig. 6 illustrates the norm of the NAC vector, CNAC , that couples the diabatic adiabatic diagonal ij crude adia- spin- states Si and Sj. At the avoided crossing points spin-diabatic batic adiabatic of the states, the NAC norm shows a sharp field- spectroscopic MCH spike, but is almost vanishing elsewhere. If adiabatic quasi- more than one multiplicity is considered, the field-free field-dressed term adiabatic is not adequate anymore, be- diabatic . cause potentials of different multiplicity might cross through each other. This situation is then called diabatic with respect to the spin multi- 3.2.1 Adiabatic (Spin-Diabatic) Basis plicities, or spin-diabatic in short. For exam- The direct output of an electronic structure ple, singlets are adiabatic among each other, calculation usually provides the eigenenergies triplets are adiabatic among each other but sin- and eigenfunctions of the electronic Hamilto- glets are diabatic with respect to triplets. How- nian. In many cases, only one spin multi- ever, also the diabatic basis (see Fig. 6(a) and plicity is calculated. If this procedure is re- also below) qualifies as spin-diabatic. Due to peated along a nuclear coordinate, potential this nomenclature issue, which even gets ex- curves result that are termed adiabatic. Adi- perts confused sometimes, we refer to this ba- abatic means ”not going through” (from greek sis as MCH (Molecular Coulomb Hamiltonian)

16 Figure 6: (a) Example of three potential energy curves ordered by their character along with respective potential couplings between different states shown by dashed lines. (b) Two singlets NAC (Ei and Ej) and one triplet state (Ek) including coupling values (with vectorial properties, Cij , shown by their norm) in the the adiabatic basis, in which the triplet state crosses singlet states. (c) The diagonal, or spin-adiabatic, basis, in which all states are ordered by their energy and are spin-mixed. Kinetic couplings are shown by their norm. Note that the ground state is not shown. because it is obtained from the eigenfunctions prises, among other terms, a relativistic part. and eigenvalues of the non-relativistic electronic This additional part of the Hamiltonian ac- Hamiltonian, where only Coulomb interactions counts for spin-orbit effects and is proportional are considered. to the atomic charge,34,36,337,339,340 leading to As an example, a crossing of a singlet state the belief that SOCs would only be relevant and a triplet state is shown in Fig. 6(b). As it in systems with heavy atoms.341,342 Today it is is visible, the triplet components, which are de- known, that spin-orbit effects also play a cru- fined by different magnetic quantum numbers, cial role in many other molecular systems and are degenerate. The states are coupled by SOCs are important for intersystem crossing between 37,343–345 (denoted as CSOCij ), which are usually obtained states of different spin multiplicities. as smooth potential couplings with standard The states in the MCH basis can also be cou- quantum chemistry programs:28,242,334 pled via external electric-magnetic fields, e.g., by sunlight or a laser. The corresponding cou- ˆ SO CSOCjk = hΨj | H | Ψki. (8) plings stem from the transition dipole moments multiplied with the electric field. Since the These couplings are single real-valued or effect of the field is not included in the po- 34,336 complex-valued properties. Whether they tentials but as off-diagonal potential couplings, are complex or not depends on the electronic the MCH basis is also called field-free.331–333,346 structure program employed, but they can be However, also the diabatic basis qualifies as 34,36,242,337 converted into each other. field-free. Hˆ SO in eq. 8 is the spin-orbit Hamilton op- erator, which describes the relativistic effect 3.2.2 Diabatic Basis due to interactions of the electron-spin with the orbital angular momentum, allowing states In the diabatic basis, the electronic wave func- of different spin-multiplicities to couple.337–339 tion is not parametrically dependent on the nu- Note that also SOCs between different states clear coordinates. Note that such a strictly dia- of the same multiplicity exist except for sin- batic basis for polyatomic systems does not ex- glets. No exact expression on how to include ist in practice and only approximated, so called, relativistic effects into the many-body equa- quasi-diabatic, PESs can be fit. In literature, tions has been found, yet. Among the most quasi-diabatic PESs are most often referred to popular approximations used is the Breit equa- as diabatic ones, so we will also use this nota- tion,340 applying an adapted Hamiltonian in- tion here. Further, diabatic potentials usually stead of the electronic Hamiltonian, which com- need to be determined from adiabatic potentials

17 and are not unique, i.e., they rely on the method discussed in detail in section 4.2). In the case and the reference point, which is chosen in the of two states, U, is a rotation matrix: adiabatic basis to fit diabatic potentials.228,242  cosθ(R) −sinθ(R)  An example of a system in the diabatic ba- U = (10) sis as given in panel (a) of Figure 6 and com- sinθ(R) cosθ(R) monly used notations can be found in Table 1 in and is dependent on the rotation angle, θ. Ac- the first column. In regions, where an avoided cordingly, the peaky NACs, which are obtained crossing is present in the adiabatic basis, the as derivative couplings (also called kinetic cou- coupled diabatic potential energy curves cross. plings) in the MCH basis, are converted to Since the electronic wave function of a state is smooth potential couplings in the diabatic ba- ideally independent of the nuclear coordinates, sis. The smooth SOCs from the MCH basis its character is conserved. Consequently the become even smoother (ideally constant) in the states are labeled according to their character diabatic basis. and multiplicity, e.g., as 1ππ∗ or according to While one can straightforwardly apply diago- symmetry labels. Similar to the character, also nalization to convert diabatic PESs to adiabatic spectroscopically important quantities like the PESs (and similarly adiabatic PESs to diago- dipole moment are mostly conserved or vary nal PESs), a dilemma arises when one wants to smoothly along the nuclear coordinates. There- take the inverse way to obtain diabatic PESs fore, spectroscopic experiments can easily be in- from adiabatic ones (and similarly adiabatic terpreted when using the diabatic basis, which PESs from diagonal ones). In fact, finding dia- is thus sometimes also called spectroscopic ba- batic PESs is highly complex and most often re- sis. Note that sometimes labels like S , etc. 1 quires expert knowledge. Up to date, only small are used also when referring to the diabatic ba- molecules could be represented with accurate sis, especially in experimental papers when an diabatic potentials and developing a method to identification of the wave function’s character automatically generate diabatic PESs remains has not been carried out and only one geome- an active field of research. Existing methods to try is considered. However, at a different ge- obtain diabatic potentials require human input ometry, the energetic order of the states might and are mostly applicable to small systems and have changed such that a state previously la- certain reaction coordinates. Early pioneering beled as S might now be lower in energy than 2 works can be found in refs.228,347 Today, a lot a state previously labeled as S . Furthermore, 1 more variants exist. Examples are the propaga- this labeling scheme in the diabatic basis can tion diabatization procedure,348 diabatization lead to confusion with the labels from the MCH by localization,349 Procrustes diabatization239 basis, and we suggest to reserve it only for the or diabatization by ansatz.140,350 Further, meth- MCH basis. ods can be based on couplings or other prop- Due to the mostly conserved characters and erties,351–354 configuration uniformity,355 block- the crossing of states, diabatic potentials are diagonalization,356,357 CI vectors358 or (partly) smooth functions of the nuclear coordinates, on ML.140,141,350,359–362 in contrast to adiabatic potentials. A diabatic PES is thus highly favorable for several numer- ical applications including ML. 3.2.3 Diagonal Basis The MCH and diabatic bases can be intercon- As the name indicates, the diagonal basis can verted by a unitary transformation be obtained by a diagonalization from the MCH

MCH diab or diabatic bases. In this case, a strictly adi- Ψ (r, R) = U(R)Ψ (r, R) (9) abatic picture is obtained, where states never cross.242 Accordingly, the concept of multiplic- with a unitary matrix, U, that is determined ity for a single state is lost because the state up to an arbitrary sign (as a result of the arbi- might be of singlet character in one region and trary sign of the wave function, which will be of triplet character in another region. There-

18 fore, the basis is also called spin-mixed or spin- tion has to be solved:240 adiabatic.36,336,363 The states are strictly or- ∂Ψ(r, R, t) ˆ dered by energy and can be labeled simply ih¯ = Hel(r, R)Ψ(r, R, t). (11) with numbers (see Fig. 6(c)). The resulting ∂t wave functions are eigenfunctions of the rel- From a technical point of view, a sequence of ativistic electronic Hamiltonian.36,242,344 These time steps is computed, where in every step the eigenfunctions as well as the eigenenergies can electronic problem is solved to yield potentials, be also obtained directly with e.g. relativis- which determine the forces acting on the nuclei tic two-component or four-component calcula- such that the nuclear equations of motion can tions,364 instead of via diagonalization. be solved for the current time step. In this basis, the effect of the SOCs are incor- Ideally, the nuclei are treated quantum me- porated into the PESs to a large extent. What chanically. In this case, the PESs are usually remains are localized kinetic couplings, which computed in advance and either interpolated or are similar in nature to the NACs in the MCH stored on a grid for later use. The hope is that basis. An example is given in Fig. 6(c). The ML can improve the interpolation of potentials parts of the potentials that correspond to the drastically. Such global PESs are needed be- different triplet components in the MCH basis cause a wave function is employed for the nu- are split energetically in the diagonal basis. In clei, which extends over a range of nuclear co- the case of small SOCs, the diagonal potentials ordinates at the same time (see Fig. 7(a)). An look similar to the MCH potentials. However, overview over corresponding dynamics methods if the SOCs are strong, potentials that are de- is given in section 3.3.1. generate in the MCH basis can be easily shifted The nuclear dynamics can also be approxi- apart by 1 eV in the diagonal basis. Such split- mated classically while quantum potentials are tings are then also experimentally observable, used, i.e., mixed quantum classical dynamics and the diagonal basis yields a more intuitive (MQCD) simulations are carried out. Such interpretation of these experiments.45,365,366 methods is discussed in section 3.3.2. Since the As mentioned above, the states in the MCH classical nuclear trajectories are defined only at basis can also be coupled via electromagnetic one nuclear geometry at a time (see Fig. 7(b)), fields. A diagonalization of the potential ma- on-the-fly calculations of the potential energies trix then yields so-called field-dressed states are possible. An on-the-fly scheme is compu- or light-induced potentials, which can also be tationally advantageous, if the number of vis- termed field-adiabatic.331,346,367–369 Since the ited geometries during the dynamics is smaller fields are usually time-dependent, the most im- than the number of points needed to represent portant axis along which the potentials in this the conformational space on a grid or via inter- field-dressed basis need to be plotted is time.346 polation.26,28,242,313,314,344,370 No fitting of PESs In principle, all these bases are equivalent but is necessary in an on-the-fly approach but fit- only if an infinite number of terms is considered ted PESs can still be used as an alternative. in eq. (6). In practice, potentials represented in Since ML approaches provide such interpolated different bases have different advantages for dy- potentials, the amount of training points gener- namics simulations, especially in combination ated with quantum chemistry must be less than with different approximations made in the dif- the number of points needed in an on-the-fly ap- ferent dynamics methods as outlined below. proach in order to be advantageous. This de- mand is satisfied, e.g., for long time scales or if 3.3 Excited-State Dynamics Sim- many trajectories are necessary. ulations In the following, we will shortly discuss the different types of nuclear motion and the oppor- In order to investigate the temporal evolution tunities of ML models to enhance the respective of an isolated molecular system in the excited dynamics simulations. states, the time-dependent Schrödinger equa-

19 20 years, (modified) Shepard interpolation is used to fit diabatic potentials.149,377–380 No- tably, the grow algorithm149 can be used to ef- ficiently generate the database of points upon which the interpolation is based. However, it is clearly desirable to treat larger systems, and ML models like neural networks (NNs) promise higher performance or more flexibility in such cases.141,144,145,147,348,359–362 More recently, on-the-fly methods address- ing quantum dynamics have been devel- oped.143,381–383 They mostly rely on a combina- Figure 7: Excited-state dynamics can be tion of Gaussians to represent the nuclear wave treated with (a) quantum approaches, where function.26 For example, the variational multi- wave functions are used for the nuclei, or (b) configuration Gaussian method (dd-vMCG)384 classical approaches, based on trajectories. offers a variational and thus accurate solu- tion for the equations of motion. Also full multiple spawning46,371,385 can be regarded as 3.3.1 Quantum Nuclear Dynamics fully quantum mechanically by describing the The computational cost of an exact nuclear dy- wave function with a number of time-dependent namics simulation scales exponentially with the Gaussian functions, that follow classical trajec- nuclear degrees of freedom. Hence simulations tories with quantum mechanically determined are limited to small systems, typically contain- time-dependent coefficients. In its more afford- ing less than 5 atoms.27,34,371 Still, the calcu- able ab-initio multiple spawning variant, more lation of the PESs of the molecule can be a approximations are introduced such that the rather expensive part of the whole scheme and results sometimes draw near the classical solu- the use of ML algorithms is advisable even for tions.386,387 Further related methods exist, like such small systems. the ab-initio multiple cloning method,388 or the To treat larger systems, approximations have thawed Gaussian approximation.389 to be invoked. A prominent approach that Another class of dynamics methods are semi- can be converged to the exact solution is the classical approaches, which allow the inclu- multi-configurational time-dependent Hartree sion of quantum effects in the classical dy- (MCTDH) approach.49,372–374 Its high efficiency namics of nuclei, such as quantum mechan- stems from the use of time-dependent basis ical tunnelling or coherence.390 Note that functions to represent the nuclear wave func- these methods, where the nuclear dynamics tions. Nonetheless, the computations are com- is treated semi-classically, should not be con- putationally costly and the nuclear degrees of fused with the MQCD approaches (see below) freedom are often reduced to only a few im- that are also often termed semi-classical (be- portant key coordinates,228,375 where classical cause the nuclei are treated classically and the simulations can help identifying the latter.376 electrons quantum-mechanically). The semi- Whether quantum dynamics of such reduced- classical dynamics methods range from the ini- dimensionality models are better than using tial value representation,391,392 adapted with classical dynamics of a full-dimensional sys- the Zhu-Nakamura approach leading to the tem is still under debate and probably de- Zhu-Nakamura-Herman-Kluk initial value rep- pends on the system. The potentials need resentation,393 to path integral approaches.394 to be presented to the algorithm in the di- The path integral formalism is especially in- abatic basis, mostly due to numerical stabil- teresting when the quantum and classical de- ity (e.g., smooth couplings are easier to inte- grees of freedom should be coupled in a dynam- grate than singular ones). Since more than ically consistent manner. By using so-called

20 ring-polymers, i.e., replica of the original clas- tum potentials, hence only one state is consid- sical system, a deviation of the nuclear dynam- ered to be active, but transitions between dif- ics from the classical path can be obtained and ferent states are allowed.406 the time evolution of a system including nuclear Different approaches exist to determine the quantum effects can be investigated. However, probability of such a transition, also called hop ring-polymer dynamics suffer from high com- or jump in surface hopping methods. To this putational efforts as a consequence of the large aim, different quantities are needed that are number of replica required. Accelerated formal- commonly provided in the MCH basis, as it ism exist, which are for example implemented in is the direct outcome of a quantum chemical the Python wrapper i-PI,395,396 which allows to simulation. One of the first implementations interface path-integral methods with programs to compute the hopping probability is based that provide PESs, but are mostly dedicated to on the Landau-Zener formalism.407,408 Based on the electronic ground state. Up to date, only a the Landau-Zener formula, the potential en- few implementations of semi-classical methods ergy differences are used to determine the hop- in atomistic simulation software are available. ping probability. No information about cou- Compared to classical mechanics, the compu- plings is required, which implies that the ap- tational costs increase by a factor of about 10 proach must fail for states that do not couple to 100.390,397,398 but lie close in energy. Very similar to this approach is the Zhu-Nakamura theory.409–412 3.3.2 Mixed Quantum-Classical Molec- Also here, the computation of couplings is ular Dynamics omitted and only information about PESs is used. Among the mostly used hopping algo- While semi-classical methods are promising to rithm is Tully’s fewest switches algorithm,403 simulate the dynamics of molecular systems which is valid for many cases and based on containing up to tens of atoms highly ac- the NACs between different PESs. An exten- curately, the study of larger systems is still sion to other couplings is provided e.g. in the dominated by computationally cheaper MQCD SHARC (surface hopping including arbitrary methods, where the nuclear motion is treated couplings) method.344 When couplings are con- 27,397–399 fully classically. In contrast to quan- sidered, an internal transformation from the tum dynamics, the motion of the nuclei can be MCH basis to the diagonal basis is most ad- computed very fast using classical mechanics, vantageous because the localized couplings of and the computation of the PESs, on which the the diagonal picture precisely indicate, where nuclei are assumed to move, remains the time the few switches of the fewest switches approach limiting step. In this sense, ML models have a should take place. In cases, where the PESs are huge potential to enhance MQCD simulations fit in advance, either with ML models or other by providing the electronic PESs and enabling types of analytical functions, the use of a di- the investigation of reactions that are not fea- abatic basis is favorable (because of the Berry 13,400–402 sible with conventional approaches. In phase, see below) but should be transformed fact, most studies that describe photochemistry to the diagonal picture for the calculation of with ML up to date aim to replace the quan- hopping probabilities. Other flavors to account tum chemical calculation of the PESs in MQCD for transitions exist. However, they have not approaches. been applied in simulations with ML algorithms The most popular MQCD method is trajec- yet. Interested readers are therefore referred to 403–405 tory surface hopping, schematically rep- refs36,48,242,314,344,403,410,413–417 for further infor- resented in Figure 7(b). A manifold of indepen- mation. dent trajectories is required to obtain statisti- The bottleneck of approaches that require cally relevant results and to mimic the extended NACs is that the computation of the couplings nuclear wave functions. For a single trajectory, remains one of the most expensive part of a the nuclei move classically on one of the quan- quantum chemical calculation. The computa-

21 tional effort to compute a NAC vector is com- for ML. The computed observables should then parable to that of a force calculation. However, be directly compared to experiments. more NACs are present than there are forces, i.e. NS ×(NS −1)/2 NACs need to be computed, 3.4 Dipole Moments and Spectra whereas NS forces are needed (respectively with entries for the Cartesian coordinates of each nu- An important property for comparing experi- cleus). Note that in case of fitted PESs with ment and theory is the dipole moment. The ML, all of these vectors have to be computed permanent dipole moment of the ground state for each data point. Conventional approaches is a frequent target of studies with ML.109,424–435 with an ab-initio on-the-fly evaluation of the The permanent dipole moment, µi (or µii) , of PESs can make use of the fact, that only one a state i can be obtained via the dipole mo- active state needs to be considered at a certain ment operator (see eq. (12) below) or as the time step. Many MD programs therefore only sum of partial charges, qa,i of atom a in state require a computation of the forces of the active i, and the vector that describes the distance of state and the respective couplings arising from the position of atom a to the center of mass of PNA this state. the molecule, rα: µi = a = qa,irα. It can Note that despite the benefits of MQCD sim- be used for the computation of infrared spectra ulations, they obey micro-reversibility only ap- with MD simulations. The spectrum is then proximately418 and effects due to coherences or obtained as the Fourier transform of the time tunneling necessitate additional considerations auto-correlation function of the time derivative as a consequence of the classical treatment of of the dipole moment.436 nuclear motion.419 In contrast to the ground state, excited-state A more approximate approach is the Ehren- simulations often make use of the transition fest dynamics method, also referred to as mean- dipole moments, which are computed from the field trajectory method. It is often used for dipole moment operator within many quantum large systems and also frequently in material chemistry programs: science.178,189 The Ehrenfest method is based on the approximation that nuclei move classically µij = hΨi | µˆ | Ψji. (12) on an average potential, rather than switching from one specific state to another.314,420,421 Due The ground state dipole moment can differ to the treatment of each electronic state sepa- strongly from those in the excited states, due to a frequency shift and altered electron distri- rately, surface hopping methods allow the accu- 437 rate bifurcation into different reaction channels, bution upon light-excitation. while such effects are neglected in a mean-field Transition and permanent dipole moments treatment of PESs. can be fit with the charge model of ref. 109, The main limitation of MQCD approaches are where point charges are never learned directly, but instead are inferred as latent variables by an the expensive evaluation of ab-initio potentials, 435 which allows dynamics simulations only for up NN dipole model making use of rα. Notice- to a couple of picoseconds. In addition, rare re- ably, the computation of absolute values of per- action channels are hardly explored as a result manent and transition dipole moments is very of usually bad statistics.36,422,423 In this sense, challenging even when highly accurate quantum chemistry methods are employed and experi- MQCD simulations offer a perfect place for ML 93,438 to enter this field of research and advance it sig- mental values are hardly reproduced. How- nificantly. The fast evaluation of the ML PESs ever, also experimental studies provide absolute can help to explore different reaction channels values only in few cases. Most computational and to obtain accurate reaction kinetics. Ob- studies therefore do not aim to reproduce the servables and macroscopic properties can be absolute values of transition dipole moments computed directly or with post-processing as but rather use relative values to obtain reason- well as analysis runs, and offer another fulcrum ably accurate absorption spectra, which can be

22 compared to experiments.56,240,242,439–441 Since 4.1 Choosing the Right Refer- many molecules absorb in the UV, the terms ence Method for Excited- UV spectra and absorption spectra are often State Data used interchangeably. However, absorption can take place in many regions of the electromag- Many existing training sets for ML in quantum netic spectrum, including, e.g., X-rays, where chemistry are based on DFT.101,103,110,443–446 rather core electrons than valence electrons are The ease of use and low computational costs excited.31 of DFT-based methods make them suitable to As already mentioned shortly, absorption treat large systems with acceptable accuracy. spectra can be obtained from a calculation of In fact, DFT is the workhorse of many stud- excited-state energies and oscillator strengths, ies solving ground-state problems. In con- which are proportional to the squared transi- trast, TDDFT has not yet managed to equal tion dipole moments. Noticeably, the transition DFT for the treatment of excited-state prob- dipole moment is only defined up to an arbi- lems. Consequently, training sets for the ex- trary sign as a result of the arbitrary phase of cited states are less frequently computed with the wave function (see section 4.2). To circum- TDDFT91,95,96,160,327,447 and rely most often on vent this ill-definition, oscillator strengths or multi-reference methods. Examples of applied the lengths of dipole vectors can be fitted with methods are CASSCF13,137–139,143,144,147,158 or ML. However, this workaround can be problem- MR-CI schemes,12–14,90,92–94,140,142,448–453 where atic if explicit field-dipole interactions should be the latter method is more expensive than the considered with ML models. former and therefore limited to describe small systems. In general, the computation of excited-state 4 Data Sets for Excited PESs is much more expensive than the com- States putation of the ground state potential of the same molecule. Not only highly accurate ab- The basis of any successful ML model is a com- initio methods have to be applied for many prehensive and accurate training set that can systems, but also forces and couplings are re- describe the required conformational space of a quired for the considered states. A high den- molecule comprehensively and accurately with sity of electronic states present in a molecular as little noise as possible.442 While electronic system can thus increase the costs of a calcu- structure theory for ground state problems is lation considerably. In this regard, an active, almost free of noise, the same cannot be said so efficient and meaningful training set generation easily for problems in the excited states. "Bad is indispensable, especially when photodynam- points with abrupt changes"14 within ab-initio ics simulations are the target of a study. calculations for the excited states are frequently Keeping in mind, that the quality of the refer- observed, which can occur even far away from ence data confines the quality of an ML model, any critical point of the PESs and are difficult several key questions can be identified when de- to detect.13,14,92 The amount of noise in the ref- signing a study based on ML potentials. We erence data does not only depend on the chosen believe the following questions to be important method (and in case of multi-reference methods for the selection of a suitable reference method: on the selected active space), but also on the 1) What is the goal of an ML model and what number of electronic states considered and the properties must it predict in order to benefit photochemistry of the molecule under investi- from the advantages that ML can offer? Are gation. only energy gaps of different electronic states to the electronic ground state necessary or are gaps between other states and couplings be- tween them also relevant? Especially the de- scription of couplings requires further consid-

23 eration, as they cannot be calculated with all tions remain the bottleneck even when using quantum chemistry methods and additionally ML. face the problem of random sign jumps along In addition to the aforementioned intricacies different reaction coordinates.90,92,454 to build up a meaningful, yet accurate train- 2) How many excited states are relevant and ing set for the excited-states, the process is which method is computationally affordable to further complicated by the arbitrary phase of treat the amount of states required? A compar- the wave function. As a consequence, excited- ison with experiment and the computation of state properties resulting from two different vertical excitation spectra with reference meth- electronic states, such as transition dipole mo- ods can help to obtain an answer to this ques- ments or couplings between different electronic tion. states,13,14,90,92,93,454 are not uniquely defined 3) How large is the system under investiga- and cannot simply be fitted with conventional tion and how complex are the excited state ML models. Either an additional data prepro- processes that are considered to be important? cessing or an adaption of the learning algorithm This question is important in order to identify has to be incorporated to render data learnable if single reference methods like LR-TDDFT or with ML models. ADC(2) make sense for certain reactions that might occur. While large and flexible molecules 4.2 Phase of the Wave Function with a lot of energetically close-lying states can give rise to a multifaceted photochem- In contrast to ground state properties, excited- istry including dissociation, homolytic bond- state properties such as transition dipole mo- breaking, and bond-formation, the dynamics of ments, NACs or SOCs arise from two differ- rigid molecules might only be dominated by one ent electronic states. As a consequence of the main reaction channel and lose the additional arbitrary phase of the wave function of each energy in form of molecular vibrations. The electronic state, properties resulting from two complexity of the excited-state processes can different states carry an arbitrary sign, which help to estimate the number of necessary data makes them generally double-valued. In case points to describe the relevant configurational of vectorial properties, such as dipole moments space of the molecule. or coupling vectors, the whole vector can be In case multi-reference methods are neces- multiplied by +1 or -1 and is still a valid so- sary to describe many different excited-state lution. Similarly, single valued properties, such processes of a molecule, the training set gen- as SOCs obtained from electronic structure pro- eration can become infeasible. For example, grams, can be multiplied by +1 or -1 and 356 data points were computed for the 15-atom are equally correct. This additional complex- cyclopentoxy molecule with MR-CISD(5,3)/cc- ity prohibits that conventional ML algorithms pVD(T)Z.94 Respective calculations comprised learn such raw data of quantum chemistry and 19,302,445 configuration state functions and hampers the training process to find a proper one reaction coordinate could be fitted in the relation between a molecular geometry and the diabatic basis. We also ran into a similar prob- excited-state property.92,454 lem when fitting the excited states of the amino A one-dimensional example of this problem is acid tyrosine containing 24 atoms, which also illustrated for the NAC (exemplified using one requires a multi-reference treatment. The size single value along the reaction coordinate) that of the active space and the number of states couples an excited singlet state, Si, and a sec- needed for an accurate description made multi- ond excited singlet state, Sj, in Figure 8. A reference methods like CASSCF or CASPT2 positively signed function of atomic coordinates computationally too expensive, see Fig. 5. In is shown by dashed blue lines with a cusp at these cases, the computation of an ample train- the point at which the two singlet states are ing set is far too expensive with multi-reference degenerate. Such a smooth function (besides methods and the quantum chemistry calcula- the sharp spike at the conical intersection) is

24 highly desirable when fitting with ML models It is worth mentioning at this point that also is aimed for. It is worth mentioning that a another kind of phase exists that cannot be consistent negative sign (light-blue dashed line) eliminated in the aforementioned way. It is along this reaction coordinate is equally correct called Berry phase or geometric phase. After and that it is desirable to seek for one global performing a loop in space around a conical in- sign. However, the direct output of a quantum tersection and returning to the original point, chemistry program along this reaction coordi- a change in the phase of the wave function of nate looks more similar to the dashed magenta π can be observed, i.e., the same point is only line in-between the blue curves. As one can reached after two loops around the conical in- imagine, no proper training can be guaranteed tersection. Neglecting this effect can lead to with these inconsistent data. Note that exist- false transition probabilities, depending on the ing MD programs for the excited states usually dynamics method and the system. While in track such phase jumps within electronic wave most cases in MQCD the Berry phase can be functions in order to account for nonadiabatic safely neglected, this is not possible in quan- transitions correctly.242 tum dynamics simulations. A diabatic basis is advantageous in this case, because the Berry phase is absent in this picture. However, the Berry phase has to be kept in mind, when fit- ting diabatic potentials.456–460

4.2.1 Phase Correction of Adiabatic Data First ML studies on dynamics in the adiabatic basis omitted a preprocessing and were un- able to reproduce reference results based on ML alone,138 or avoided the phase problem by using the Zhu-Nakamura method.137,139 Ev- Figure 8: NAC value between singlet state Si idently, potentials and forces can be learned and Sj in the MCH basis. A consistent sign along the reaction path of couplings is shown by with conventional ML approaches but adapta- blue dashed lines. The direct output of a quan- tions or a preprocessing of data is necessary to tum chemical calculation is shown by a magenta learn coupling elements or transition dipole mo- line. ments. Independent of the purpose – the fitting of adiabatic quantities92,454 or the diabatization of adiabatic data with property-dependent dia- The idea of phase tracking can also be ap- batization schemes14 – the adiabatic data has to plied in ML in order to thwart the problems be corrected to remove the arbitrary sign jumps due to the arbitrariness within coupling or that are due to the arbitrary phase of the wave dipole elements. Some algorithms have been function. Several ways for these corrections ex- developed to remove the arbitrary sign jumps ist, which have been shown to work well for and provide smooth functions of atomic coordi- different excited-state problems. nates.13,14,92,455 Noticeably, the properties ob- One possibility is to preprocess data accord- tained after a transformation to the diabatic ing to the wave function overlap – betweem the basis are already smoothly varying functions of wave functions from a geometry of interest and atomic coordinates.336 However, the challenges a reference geometry – for each electronic state. arising due to the arbitrary phase of the wave This process is termed phase correction242,454 function still persist, because the inconsisten- and has been applied by us in order to gen- cies within adiabatic properties have to be re- erate a training set for three singlet states of moved in order to make the diabatization pro- CH NH+ 92 and 2 singlet and 2 triplet states cess feasible.14,90 2 2

25 13 13,92,93 of CSH2. SOCs, NACs, and transition phase vectors, p0 to pn−1: dipole moments92,93 could be fitted in the adi- n−1 abatic basis with deep NNs and kernel ridge Y p = p . (14) regression (KRR).13,92,93 Very recently, Zhang α et al.95 applied this procedure to describe tran- α=0 sition dipole moments of N-methylacetamide. Intruder states prohibit a proper tracking be- The wave function overlap matrix, S, with cause their wave function is absent at the ear- size N ×N , is computed between two molecu- S S lier geometries. Hence, a phase correction may lar geometries α and β:461 be rendered infeasible for systems with a high density of states. S = hΨα | Ψβi. (13) In order to obtain the correct phase, more In many cases along a given reaction path, the states can be included in the simulations, which off-diagonal elements of the overlap matrix are however increases the computational cost. A very close to zero and the diagonal elements are solution is to take many electronic states into very close to +1 or -1, indicating whether the account only close to the reference geometry. phase of a state has changed along this path The amount of states can then be reduced along or not. Whenever a new state enters along a given reaction coordinate and relevant states the reaction path or adiabatic states switch can be disentangled from irrelevant ones. Fur- their character, which is common after passing ther, it makes sense to save the already phase- through a conical intersection for example, the corrected wave functions of several geometries off-diagonal elements provide the relevant phase in addition to the reference geometry. When- information instead of the diagonal elements. ever a new data point should be included into Taking all these effects into account, a phase the training set, the distance to each saved data vector, p, can be derived for each given molec- point can be computed in order to find the clos- ular geometry. A property resulting from elec- est available structure and reduce the amount tronic state i and j has to be multiplied by the of interpolation steps.92,400 corresponding phase factors of these states.92 This problem has also been recognized by An advantage of this algorithm is that it does Robertson et al.358 for a diabatization process, not require any manual fitting of data. How- where a sufficiently large vector space of the ever, this procedure has to be carried out for CAS wave function is required for proper di- every data point included in the training set abatization. The overlaps of electronic states with respect to one pre-defined reference wave can be maximized by rotation of CI vectors of function. This reference wave function can be CAS wave function states. A similar version to for example the wave function of the ground- use the information of CI vectors for diabati- state equilibrium structure of the molecule and zation was applied by Williams et al.,140 who needs to be identified to guarantee an almost used NNs to assist the diabatization process of globally consistent sign of elements. During a adiabatic NO3 potentials. photo-initiated simulation, it is common that Another way to correct the sign of data points geometries quickly start to differ from the refer- was carried out by Guan et al.,14 who fitted di- ence geometry. The wave function overlap then abatic 1,21A PESs and dipole moment surfaces tends to zero and cannot provide information of NH3 from MR-CISD/aug-cc-pVTZ data with about the correct sign of a certain electronic NNs. The diabatic PESs were taken from a pre- state. In this case, the phase must be propa- vious study and obtained with the Zhu-Yarkony gated from the reference geometry on with n diabatization procedure.462–464 By diagonaliza- interpolation steps. The phase vector applica- tion, the rotation matrix defined in eq 10 could ble for the correction of the data point to be be obtained, which connects the diabatic and included in the training set is then obtained the adiabatic basis (see eq. (9)). The adiabatic by multiplication with all previously obtained dipole moments, µMCH , could then be trans-

26 formed into the diabatic basis using the unitary manually for this purpose. The diabatic SOCs matrix, U: were then obtained as a linear combination of the adiabatic SOCs by applying the same ro- diab MCH † µ = Uµ U . (15) tation matrix as for the energies. One separate NN function was used to fit each coupling value As the unitary matrix U is only defined up and electronic state separately. to an arbitrary sign, the signs of the diabatic It becomes clear that only a small number dipole moments have to be corrected in order of works on this topic exist. At the moment, to provide a consistent diabatic dipole moment many problems remain unsolved for generating surface. This correction has been done with a a training set that properly accounts for both 455 so-called cluster growing algorithm. types of phases, the arbitrary phase and the The cluster growing algorithm requires an ini- Berry phase, and is applicable for large sys- tial set of phase corrected data points. In this tems with many states. An automatic phase work, 347 data points were adjusted manually correction procedure without the need of man- for this purpose. Subsequently, a Gaussian pro- ual input would be very advantageous, espe- 465 cess regression (GPR) model was fitted to cially when larger and more flexible systems are these data points. The signs of the rest of the treated. Further developments are needed. data points to be corrected were then adjusted with the GPR model. Several iterations were 4.2.2 ML-Based Internal Phase Correc- carried out, where each iteration aims for the in- tion clusion of close-lying points to the cluster, lead- ing to the name "cluster growing" algorithm.146 One step towards a routine application of ML The singularities in regions close to conical in- for photochemical studies and an easier train- tersections can make this algorithm fail. There- ing set generation with quantum chemistry is fore, data points in such regions have been re- an ML-based internal phase correction, which moved by setting a threshold. Data points with has been implemented by us into the SchNarc energy gaps lower than this threshold were ex- approach for photodynamics simulations.13 In cluded from the cluster. The regions around contrast to the phase correction algorithm to conical intersections could not be fitted as com- correct the training data, this procedure ren- prehensively as other regions of the PESs. As ders the learning of inconsistent quantum chem- another drawback, the authors note that the ical data possible. A modification of the train- initial manual fitting of the signs is a tedious ing process, termed phase-free training, is re- task, especially when larger systems and more quired for this purpose.13 dimensions are described. We implemented this training algorithm in Two of the authors also fitted diabatic PESs a combination of the deep continuous-filter of two singlet states and one triplet state as convolutional-layer NN SchNet,428,432 adapted well as the SOCs between singlets and triplets for excited states, and the MD program 90 242,344,466 of formaldehyde, CH2O, with NNs. The elec- SHARC tronic structure reference method was MR- Similar to standard training algorithms, pa- CISD/cc-pVTZ. The diabatic potentials were rameters of an ML model are optimized in or- obtained using an adapted version of the Boys der to minimize a cost function. Most fre- 351 localization. The energy differences between quently, the L1 or L2 loss functions are applied, two states are incorporated in the equations in which take the mean absolute error or mean order to remove earlier identified diabolic singu- squared error between predicted and reference larities.146 The range of π, which the rotation data into account. The phase-free training al- angle for the diabatization covers, guarantees a gorithm uses a phase-less loss function, which proper treatment of the Berry phase. The dia- includes all trained properties at once and ad- batization procedure further requires consistent ditionally removes the influence of the random transition dipole moments, which were adjusted phase switches. In this way, the computational

27 costs for the training set generation can be re- duced. εκ = Compared to the previously reported ML NAC 1 PNS PNS 1 PNA N 2 i=1 i6=j N a=1 models for photochemistry, where each state S A (20) 14,90,139 || CQC − CML · pκ · pκ ||2 was fitted independently, SchNarc is ca- NACij,a NACij,a i j pable of describing all PESs at once, includ- with 0 ≤ κ ≤ 2NS −1 ing the elements resulting from different pairs of states. This results in an overall loss function This phase-less loss procedure does not re- with several terms, where each term is weighted quire any preprocessing of training data. Quan- with a different trade-off value, t, that can be tum chemistry calculations can be directly fit- defined manually: ted with this adaption of the loss function. The power of this approach is that, once a given Lph = phase vector for a data point has been found, QC ML 2 tE || E − E || it can be directly applied to correct the arbi- QC ML 2 +tF || F − F || (16) trary signs of other properties, such as tran- +tSOC · LSOC sition dipole moments. If other properties are +tNAC · LNAC targeted, the loss function applied for NACs can be similarly used for other vectorial properties, If only energies (E) and forces (F) are fitted, and the loss function applied for SOCs can be then the loss function is equal to a linear com- used for any other single- or complex-valued el- bination of L2 loss functions for energies and ement of arbitrary sign.13 However, as a con- 13,109 forces. The parts of the SOCs and NACs sequence of the higher complexity of the loss are function, the training process is generally more L = min(| εκ |) SOC SOC (17) expensive. The computational effort required with 0 ≤ κ ≤ 2NS −1 for training can be reduced if only one type of and coupling is treated within MD simulations. In κ LNAC = min(| εNAC |) these cases, a simpler adaption of the phase-free N −1 (18) with 0 ≤ κ ≤ 2 S , loss is also applicable.13 respectively. The error for SOCs and NACs that enters the loss function is the minimum 4.3 Training Set Generation error that can be achieved when trying out all possible combinations of phases for each pair of The requirements and desirable specifications states, i.e., 2NS −1 possible solutions. The algo- for a training set can vary strongly, dependent rithm takes into account that the signs of SOCs on the type of application: When the focus of a and NACs coupling different pairs of states de- study is the investigation of the huge chemical pend on each other. space and the search for certain patterns thereof The error function containing all possible so- or the design of new molecules with targeted phase phase properties, usually the training set should be lutions for SOCs, εSOC , and NACs, εNAC , can be obtained as follows: as large as possible to cover as many molecules as possible. In the best case, the data points κ εSOC = are computed with high accuracy and this ref- 1 PNS PNS QC ML κ κ 2 2 i=1 i6=j || CSOC − CSOC · pi · pj || erence method is accurate for the excited states NS ij ij with 0 ≤ κ ≤ 2NS −1 of many different types of systems. In terms (19) of accuracy and general applicability, ab-initio methods are more suitable, as they do not require the selection of a density functional, which might be accurate for some cases, but fail for others. However, the costs and com- plexity of highly accurate multi-reference ab-

28 initio methods limit their applicability, so that applied. A large number of schemes to achieve TDDFT remains the method of choice when this goal have been proposed, which are mainly making predictions throughout chemical com- based on two different strategies: One approach pound space.150,327,467 The most widely applied is to simulate MD in the ground and excited approach to generate a training set for this pur- states with the reference method and putting pose is to start from an existing (ground-state) much effort into covering critical regions of the data base that already covers a large chemical PESs comprehensively.137–139 Structure-based space of certain types of molecules. In this way, sampling or subsequent clustering is beneficial not much effort has to be devoted into the ex- in this case.137,138,481–483 The other strategy is ploration of chemical space and structure opti- to use an active learning approach, which de- mizations to get the most stable conformations creases the number of necessary reference calcu- of different molecules. lations considerably, but is usually more time- For the purpose of ML-based excited-state dy- consuming.480 Noticeable, within ML for quan- namics simulations, things look quite different. tum chemistry, active learning often refers to Note that for photodynamics simulations, only an approach, where an initial training set is molecule-specific ML model exist until now, used to fit an ML model and this previously which can potentially develop into a universal learned information is applied to expand the excited-state force field, but much remains to training set.402 The latter approach is often car- be done to achieve this goal. Indeed, the gen- ried out with the help of MD simulations, but eralization of the excited state PESs and corre- has also recently been adapted in a trajectory- sponding couplings is expected to be a highly free way.402,484 complex task, especially due to the problematic generalization of excited states.92 A compari- 4.3.1 Basic Sampling Techniques and + son of the isoelectronic molecules CH2NH2 and Existing Databases C2H4 can serve as an example. Their conical in- tersection between the first excited singlet state To find patterns within certain groups of and the ground state is accompanied by a rota- molecules, to explore chemical space and to tion along the dihedral angle, which could lead develop new methods that can fit for exam- to very similar photo-initiated processes. How- ple different properties of molecules, such as 81 ever, higher-lying excited states are ordered the valence density used in DFT, or large 107 completely different in both molecules and ex- molecules from small building blocks, a good citation leads to completely different photody- starting point is often considered to be an al- namics.92,468–478 ready existing data base. Prominent exam- As it stands, existing ML models for photo- ples are the QM data bases, namely QM7, 424 dynamics simulations are developed to investi- QM7b, QM8, and QM9, which have been gate the photo-initiated processes of one spe- used in a large number of publications up cific molecule. to date and provide a benchmark for many 12,150,428,433,434,446,485–489 Overall, we arrive at the following wish list ML studies. Especially 424 for the training set, which has been identified the QM9 data set containing more than also for MD in the ground state:101,115,479,480 133k small organic molecular structures and 1) The training set should be as small as corresponding DFT energies, enthalpies, har- possible to keep the number of reference cal- monic frequencies, and dipole moments (to culations at a minimum. 2) At the same name only a few properties) is very popular time, the relevant conformational space of the among the scientific community and has also molecule that is required for the reaction un- been used in challenges on kaggle, where re- der investigation should be sampled compre- searcher and layperson all over the world can hensively.92,115,328,442,480 compete against each other to find the most Keeping this in mind, an efficient procedure to suitable solution to a given task. Prices up to 490 obtain relevant molecular structures has to be several thousand dollar are quite common. In

29 a similar spirit, the QM9 IPAM ML 2016 chal- fort328 and has been adapted for spectroscopy lenge requires to predict the energies of QM9 in the condensed phase as well.151 from only 100 training points within chemical The QM9 data set has further been the ba- accuracy (error of ≈0.05 eV).491 sis of a very recently constructed data set All aforementioned data bases originate from for singlet and triplet states of >13k carbene GDB data bases,492–494 and are often a subset structures, termed QMspin.12 4,000 geometries thereof. The chemical universe GDB data bases from the QM9 data set were randomly se- have been designed using molecular graphs to lected, hydrogen atoms were subtracted and sample a comprehensive space of molecular singlet and triplet states were optimized using structures for the search of new lead compounds CASSCF(2,2)/cc-pVDZ-F12 and open-shell re- in drug design.494 stricted KS-DFT with the B3LYP498,499 func- One of the first data bases available for the tional, respectively. The MR-CI method was scientific community to treat the excited states subsequently used to compute the electronic en- of molecules is most probably the QM7b495 data ergies of singlet and triplet states. This data set, that contains the excitation energies com- set has been used to investigate structural and puted with TDDFT for a total amount of >14k electronic relationships in carbenes, which are molecules with atoms C, N, O, H, S, and Cl. important intermediates in many organic reac- This data set is based on the molecular geome- tion networks.12 tries of the QM7100,494 data set plus an addi- The OE62500 data base, a benchmark data tional amount of 7211 molecules containing a set applicable for spectroscopy, is another de- chlorine atom. The excitation energies of the scent of several existing data sets, such as the first singlet state and other properties were re- QM8 and QM9 data sets. It consists of >61k computed for each optimized molecular geom- organic molecules able to form crystals includ- etry. Very similar, the QM8327 data base was ing up to 174 non-hydrogen atoms. Reported developed, based on the GDB-17 data base.496 are the orbital energies of molecules computed This data set can be used for the computa- with DFT/PBE.501 tion of vertical excitation spectra. It hence Another database, which also contains includes not only the vertical excitation ener- excited state data, is the PubChemQC gies of the first excited singlet state, but also data base.502 It contains over three million the corresponding oscillator strengths. Oscil- molecules, whose structures are reported along lator strengths are also reported in an auto- with the energies at DFT/B3LYP/6-31G* level generated data set for optoelectronic materials of theory. In addition, the excitation energies of with DFT.467 Note that the oscillator strength at least three million structures are reported for is computed from the squared transition dipole the 10 energetically lowest-lying singlet states moment, hence an arbitrary phase factor can- at TDDFT/B3LYP/6-31G* level of theory. cels out and the data does not have to be A simple strategy was carried out by Kolb et preprocessed. In addition to the TDDFT en- al.,503 who used an existing analytical PES to ergies, CCSD energies are reported, having create an ML potential: They randomly sam- enabled the development of the so-called ∆- pled data points, trained an ML model and learning approach - a powerful way to obtain added more points in regions with deviations the accuracy of highly accurate ab-initio meth- from the original PES. Other strategies have ods with only a small amount of respective been carried out mainly for the fitting of ground reference calculations. Two ML models are state potentials and for materials, which are trained in this approach, one on a less accu- however also relevant to consider for the excited rate method and another one on the difference states. One novel, suitable strategy is for ex- between the less accurate and higher sophisti- ample "de novo exploration" of PESs using a cated method.497 This scheme can also be ap- similarity measure provided by ML models.504 plied multiple times to achieve increasing ac- At least for material discovery, this method can curacy with little additional computational ef- be used to omit any additional active learn-

30 ing procedure to converge PESs. Another way the PESs were first sampled via an inexpen- to build a training set is to employ molecule- sive method and subsequently the distances generating ML models,164,505,506 such as the re- between the molecular structures were com- cently developed Gschnet.507 Alternatively, MD puted. In this way, 10,000 data points were ob- simulations with the reference method can pro- tained.138,482 ML models trained on only 1,000 vide a good starting point for training.120,486,508 data points were accurate enough to reproduce Ye et al.509 sampled 70k conformations for N - reference dynamics. This approach was com- methylacetamide via MD simulations with the pared with random sampling for the methyl OPLS force field510 within GROMACS511 for chloride molecule and was shown to reduce the subsequent UV spectra calculations. We have amount of training data needed up to 90% for applied a similar scheme to generate a train- static calculations.482,483 229 ing set of SO2 based on an LVC model. Sur- face hopping MD simulations with the SHARC 4.3.2 Active Learning method242,344,466 were carried out with the ref- erence method LVC(MR-CISD) ending up in As shown in the previous section, training sets >200k data points of different conformations with the respective equilibrium structure of a 13 large number of molecules are very powerful for of SO2. Due to the crude sampling and low cost of the reference method, no emphasis was investigating the huge chemical space or for the put on clustering the training set into a smaller, design of new molecules. However, the useful- still comprehensive set. ness of such training sets for photodynamics is 90k data points were required in an ML- rather questionable. The reason for this de- ficiency is that, especially in MD simulations based surface hopping study of CH2NH with the Zhu-Nakamura method. Reference data for in the excited states, the excess of energy car- ried by a molecule very quickly leads to con- the ground and first excited singlet state, S0 formations that are far beyond the equilibrium and S1, were generated with CASSCF(2,2)/6- 31G via ground-state and surface hopping MD structure and most likely far away from orig- simulations. The latter method was applied to inally sampled structures. The formation and sample the regions around conical intersections breaking of bonds is quite common in photo- 139 dynamics simulations and is usually only acces- between the S0 and S1 state. Similarly, Hu et al.137 sampled 200k sible from an excited, dissociative state. The data points of 6-aminopyrimidine using use of photodynamics simulations with the ref- ground-state and surface hopping MD with erence method could solve this problem, but CASSCF(10,8)/6-31G*. State-averaging over are not feasible if specific reactions occur on three singlet states was applied. In addition, a rather slow time scale or if many different 27,28,36,37,171,400 structures that led to hops between different processes take place. As previous states were used as starting points to find mini- studies have shown, inefficient sampling tech- mum energy conical intersections and clustering niques lead to a huge amount of data, which was carried out to reduce the amount of data still does not guarantee that the training set is for training. comprehensive enough for excited-state MLMD One way to select data points more efficiently simulations. In fact, ML models fail dramat- is a structure-based sampling scheme, as pro- ically in under-sampled and extrapolative re- posed for instance by Ceriotti et al. with sketch gions of the PESs. A smarter sampling tech- map,481,512,513 an algorithm for dimensionality nique is advantageous in these cases in order to reduction of atomistic MD simulations or en- efficiently identify such under-sampled regions hanced sampling simulations. Likewise, Dral et and build trustworthy ML models. al.138 applied a grid-based sampling method to Active learning, where ML ”asks” for its train- construct PESs of a model spin-boson Hamil- ing data, is one solution to create a data tonian to execute surface hopping MD with set more efficiently. An example from chem- KRR. The energetically low-lying regions of istry is the adaption of an initially generated

31 training set due to an uncertainty measure When small molecules are targeted, this ini- for ML models trained on this initial train- tial training set can already be comprehensive ing set. This concept has already been intro- to start the training of ML models and adapt duced in 1992 as query by committee514 and the training set based on an uncertainty mea- has been adapted for quantum chemistry quite sure provided by the ML models.92 In case more fast due to the required fitting and interpola- flexible and larger molecules are studied that tion of PESs for grid-based quantum dynamics give rise to a complex photochemistry and a simulations. Pioneering works by Collins and high density of states including different spin co-workers148,149,377,515 applied modified Shep- multiplicities, a small initial training set might ard interpolation to fit PESs and iteratively not be sufficient and a larger conformational adapt them in out-of-confidence regions us- space of the molecule needs to be sampled. ing the GROW algorithm.515,516 Since then, This can be done for example via Wigner sam- several sampling techniques have been devel- pling534 and also with MD simulations in the oped that are based on MD and an exten- ground state.535,536 Suitable methods are for ex- sion of data bases using interpolation moving ample umbrella sampling,537 trajectory-guided least squares,517,518 permutation invariant poly- sampling,538 enhanced sampling539 or metady- nomial fitting,519,520 and different ML models namics540 in combination with a cheap elec- for the ground state99,101,109,479,480,521–533 and tronic structure method like the semi-empirical also excited states.13,92,138 tight-binding based quantum chemistry method As active learning starts from already trained GFN2-xTB541 or existing ground-state force ML models, an initial training set has to be fields. A large amount of different geome- provided. Some strategies to provide this initial tries can be created very fast and inexpensively, reference data set will be discussed, following which then can be clustered to exclude similar strategies applied to adapt this initial training conformations of the molecule to keep the num- set. Note that all previously discussed methods ber of reference simulations at a minimum. The can be similarly applied to generate an initial selected data points for the training set can then training set. be computed with the chosen reference method, whose accuracy is targeted with ML. Addition- Initial training set In general, an initial ally, if certain reaction coordinates have been training set can be obtained in many different shown to be important in experiments or pre- ways. As photo-initiated MD simulations usu- vious studies, then it is favorable to include ally start from vertical excitation of the ground data from scans along these reaction coordi- state equilibrium geometry, this structure is nates.94,400 commonly used as the starting point and refer- As soon as meaningful ML models can be ob- ence geometry for the training set generation. tained from the initial training set, active learn- In principle, any technique can be applied to ing techniques can be applied to enlarge the set. then add conformations to obtain a preliminary What number of data points turns out to be training set. A good starting guess is to use sufficient for the initial training set is depen- normal modes of a molecule, as they are gen- dent on a lot of different factors, such as the erally important for dynamics. In two recent size and flexibility of the molecule under inves- works, we carried out scans along different nor- tigation, the number of excited electronic states mal modes and combinations thereof to sample described, and the ML model and descriptor conformations of small molecules.13,92 Normal applied.92,93 In order to give a ballpark figure, modes are also sampled for generating ANI-1 we note that we used approximately 1000 data NN PESs.113 For the excited states, it is favor- points as initial training set for small molecules able to include critical regions of the molecule in recent studies using deep multi-layer feed- in the initial training set by carrying out opti- forward NNs.13,92 mization of these geometries and including the calculations into the training set.92,137

32 Strategies for actively expanding the ferences made by the different ML models for ML ML training set The next step in active learning energies, E , forces, F , and if required also ML is to expand the initial training set by adding couplings, C . In each sampling step, the points from out-of-confidence regions. The de- variances for each predicted property are com- tection of these undersampled regions can be puted. In the present example, energies and done in many different ways, whereby most ap- ML forces are treated together as σE+F (but can proaches rely on MD simulations. also be used separately), separately from vari- Among the most popular strategy is the itera- ance of the couplings σML. If a variance exceeds 480 C tive sampling scheme of Behler, originally de- a pre-defined threshold, the ML models diverge veloped for fitting ground-state PESs. Today, it and the predictions are deemed untrustworthy. is widely used, see for example refs 101,479,542, NML refers to the number of different ML mod- and has been modified as a so-called adap- els, ζ, used for adaptive sampling: tive sampling approach.109 The latter has been adapted by us for the generation of a training σML = E+F ! set for the excited state PESs of molecules in- r 2 1 NS 1 NML  ML 92 P P EML − E + cluding couplings. The basis of almost any it- NS i NML−1 ζ=1 ζ erative or adaptive sampling scheme is a sim- s !   ML2 ilarity measure to judge whether a molecular 1 PNML 1 P3NA ML F − F a geometry can be predicted reliably with ML NML−1 ζ=1 3NA a ζ,a models or not. While kernel methods intrin- (21) sically provide a measure of similarity for each σML = C r molecular geometry, NNs do not. Therefore,  ML2 1 PNS PNS 1 PNML ML 2 i j ζ=1 Cζ − C adaptive sampling with NNs requires at least 2NS NML−1 two ML models. In case of KRR or GPR, two (22) ML models can be used as well, but are not Note that the variance is averaged over all necessarily needed. Indeed, the statistical un- states for energies and forces and over all pairs certainty estimate of the predictions remains a of states for couplings, that are described with huge advantage of GPR models.442,543 the ML models. As a variant, each state could The adaptive sampling scheme for the excited also be treated separately. However, as the dif- states is illustrated in Figure 9 and exemplified ferent electronic states are not independent of with two ML models. The whole process starts each other, a mean-treatment is assumed to be with an initial training set, which is used to advantageous.93 train the two (or more) preliminary ML mod- Each data point that is predicted with a vari- els. These models differ in their initial weights ance larger than the pre-defined threshold for or model parameters. The resulting dissimilar a given property, is recomputed with the ref- ML architectures guarantee that the ML mod- erence quantum chemistry method and added els do not predict the exact same number for a to the training set. In this way, undersampled given molecular input. The hypothesis under- or generally unknown regions of the PESs are lying this scheme is that inferences of different identified. Whenever the variance of each prop- ML models trained on the same training set will erty is within the range that is thought to be be similar to each other as long as an interpola- reliable, the mean of the inferences is forwarded tive regime is given. The inferences of the ML to the MD program to propagate the nuclei and models are inaccurate and should differ from continue MLMD simulations. The name adap- each other to a much larger extent if a molecu- tive sampling is based on the recommendation lar input lies in an unknown or under-sampled to choose a rather large threshold in the begin- region of the PESs. ning of the adaptive sampling procedure and to In order to find such regions, sampling steps adapt this threshold to smaller values as the ML are carried out, e.g., by running (excited-state) models become more accurate and robust.109 A MD simulations based on the mean of the in- first estimate for the initial value of a threshold

33 Figure 9: Adaptive sampling scheme illustrated using two ML models (blue and red blocks). The active learning procedure starts from an initial, preliminary training set (yellow), which is used to train ML models. A sampling step, e.g. a time step of an MD simulation, is executed: The ML models take the molecular geometry of the sampling step as an input and predict the energies of the considered excited states, their derivatives, and additional required photo-chemical properties. In case the predictions of the ML models are deemed to be different, quantum chemical reference calculations are carried out, ML models are retrained and the serial steps are carried out again. This procedure is executed until the desired quality of the ML PESs is attained in order to sufficiently describe the chemical problem under investigation. can be obtained from the MAE of the corre- MLMD simulations substantially. In this re- sponding ML model on the initial training set. gard, also the computational costs for the train- In principle, adaptive sampling can be carried ing set generation can be kept at a minimum. out for every property, that should be repre- Adaptive sampling was carried out success- sented with ML potentials, and is not restricted fully to generate a training set of 4,000 data + to energies, forces, and couplings. Similarly, it points of CH2NH2 containing three singlet does not need to be executed with excited-state states and couplings. ML-based surface hop- dynamics, but could also be done with ground- ping MD simulation could be carried out on state MD or any sampling method that is con- long time scales using the average of two deep sidered to be suitable. NNs. The concept of iterative sampling also As a negative side effect, this procedure proved beneficial for the long MD simulation to is generally more time-consuming than many guarantee accurate ML potentials throughout other sampling techniques, because ML mod- the production run. Here, the threshold was els have to be trained each time a new data not adapted anymore and the MD was contin- point is added to the training set. To apply ued from the current geometry after a training adaptive sampling in a more efficient way, it is cycle was completed.92 In addition, the average advantageous to execute not only one ML tra- of more NNs turned out to be more accurate jectory, but many hundred trajectories in paral- than the prediction of only one NN, which was lel, as it is usually done in MD simulations. The also shown in Ref. 109. ML models should then only be retrained, when Another quality control besides the property- all ML-based trajectories have reached an un- based one proposed by Behler can be obtained dersampled conformational region.92,109,480 De- by comparing the molecular structures at each spite the higher complexity of adaptive sam- time step as done by Dral et al.138,482 and Ce- pling compared to random sampling, it can re- riotti et al.481 A combination of a structure- duce the number of required data points for based and property-based detection of sparsely

34 sampled regions of the PESs has been done by 5.1 ML Models: Type of Regres- Zhang et al. and Guo et al.362,524,544–546 Very sor recently, an alternative approach has been ap- plied with NNs by Lin et al.402 that does not Given the vast number of ML algorithms ap- require MD simulations. It is based on the find- plied in the field of computational chemistry, ing that the negative of the squared difference one might ask which one to use or adapt for surface obtained from NNs approaches zero in photochemistry. As recent studies applying ML regions, where no data points are available.518 for quantum chemistry have shown, many pos- Therefore, new points can be computed at the sible choices of ML approaches exist and there minima of the negative squared difference sur- is no single solution. Nevertheless, a trend can faces of at least two NNs (or, equivalently, at be observed: Many studies that use ML in the local maxima of the squared difference surface). research field of quantum chemistry employ la- This method is supposed to be very efficient in belled data sets, i.e., supervised learning tech- cases, where different conformations are sepa- niques. Within supervised learning, one can rated by large energy barriers or strongly sta- distinguish between regression and classifica- bilized local minima are common. MD simula- tion. Classification aims at finding patterns and 553 tions would take a long time to overcome the at grouping data into certain clusters. Those potential barriers and reach the region of un- types of ML models are often used e.g. in spam 554,555 known molecular structures.402 filters, in medicine to diagnose diseases, The idea behind this technique is similar to or in food research, e.g. to guarantee a certain 556 previous works with GPR. A measure of confi- wine quality or origin. Examples of applied dence can be provided with GPR models that classification models in the field of computa- enables the search of regions with large vari- tional chemistry are for example support vector ance in ML predictions. In these regions, data machines, random forests or decision trees used, 557 points can be added to build up a training e.g., to classify enzymes or for the selection 70,558 set .484,547–549 Similarly, Bayesian Optimisation of an active space. Structure Search (BOSS) has been proposed for More often than classification models, regres- constructing energy landscapes of organic and sion models are applied to assist the search inorganic interfaces.550 A combination of differ- for a solution of a quantum chemical prob- ent approaches has also been applied by Häse lem. Regression is used to fit functions that et al.,160 who fitted TDDFT excited-state ener- can relate a molecular input, X, to a quan- gies of a light-harvesting system. Given a large tum chemical output, Y . The simplest rela- enough, error-free, and comprehensive data set, tion that can be assumed is linear. Although ML has the potential to determine known and many quantum chemical problems cannot be unknown (un)physical laws within the data.551 accurately described with a linear function as given in eq. 23, it can serve as a baseline model to evaluate the minimum accuracy one can ob- 5 ML Models tain.92,171,553,559,560

Besides the training set, which defines the high- Y = b + w · X (23) est possible accuracy an ML model can attain, the type of regressor and the descriptor to rep- The regression coefficients, also known as resent a molecule to the ML model play also weights, w, and biases, b, are tailored for a given important roles.552 Improper choices of regres- problem under investigation. In case of linear sors and descriptors can result in inaccurate ML regression, ordinary least squares regression models. can be applied to find these coefficients. The process of finding the optimal relation between X and Y is termed training. The coefficients are optimized by minimizing a so-called loss

35 function, L, which monitors the error between coefficients and kernel instances: the original property, Y QC , and the predicted N ML XM property by the ML model, Y , with respect Y ML(X ) = w K(X ,X ). (26) to the training instances. Most often, the L α β α β 1 β loss or the L2 loss is used as an indicator for the training convergence. The L1 monitors the The size of the kernel matrix is dependent on mean average error (MAE) and the L2 loss the the number of training points and hence the mean squared error (MSE) of predictions: depth of the model is inherently linked to the size of the training set, which is why they are NM 553,562 1 X  ML QC  called ”non-parametric”. L2 = Yβ − Yβ . (24) NM An advantage of kernel methods is that they β mainly contain two hyperparameters, i.e., in- ternal model parameters, which need to be op- The Greek letter β runs over all molecules, NM , inside the training set. In principle, any error timized for proper training. Most important estimate can be used to train an ML model and are the width of the non-linear kernel func- find suitable regression coefficients. tion, σ, and the regularization. The latter is An example specifically developed for excited- used to prevent the model from overfitting – the state problems is the aforementioned phase-less case when the model fits training data includ- loss (see section 4.2.2).13 Such adapted loss ing noise almost exactly and fails to accurately functions and also conventional ones are em- predict data points not included in the training ployed in different types of ML models. In the set but stemming from an interpolative regime. following, we focus on the two most widely used As quantum chemical data is most often noise- models for the description of the excited states: free, the regularization term is usually small. Kernel methods and NNs. As the optimization of hyperparameters is of- ten a tedious task, kernel methods with their Kernel methods Kernel methods561 are few hyperparameters are easier to use than, e.g., based on a similarity measure between data NNs with many hyperparameters. Nonetheless, points. Examples are KRR or GPR, which go kernel methods can provide almost exact so- 125 beyond linear regression by applying the ker- lutions of problems under investigation. A nel trick and ridge regression. Ridge regression drawback is, however, that the inversion of the is used to find the weights, which differs from kernel matrix can become expensive and even linear regression by a regularization term, λ: be rendered infeasible on current computers due to increasing memory requirements with in- w = (K + λ1)−1Y QC (25) creasing training set size.93 Further, kernel methods are usually defined to YQC refers to the training data and K to the only map an input to a single output. There- kernel matrix. fore, they can treat only one electronic state The kernel trick makes it possible to apply at a time in standard implementations and, ridge regression to non-linearly separable data thus, can be referred to as single-state mod- by mapping them into a higher-dimensional fea- els. A single-state treatment requires a sep- ture space, in which the data points are lin- arate ML model for each electronic state or early separable. Therefore, a kernel function, for each property resulting of a pair of states, k, e.g. a Gaussian or Laplacian, is placed on whereas a multi-state ML model describes all each compound to measure the distance to all electronic states and properties resulting from of the other compounds in the training set. The different pairs of states at once.93,400 Hence kernel function defines the non-linearity of the in their standard implementation, the treat- model. A property of a query compound, α, can ment of several excited states necessitates the be obtained as the weighted sum of regression use of several kernel models, which is com- monly done in the research field of quantum

36 chemistry.137,138,147,563,564 The description of Due to the highly flexible functional form of forces is possible for the ground state or a NNs, highly complex relationships can be fit, single excited state and is implemented, e.g., but an analytical solution to find the weights is in the QML toolkit using KRR and the Faber- not available (in contrast to KRR). A numerical Christensen-Huang-Lilienfeld (FCHL) repre- solution can be obtained with stochastic gradi- sentation,433 in the symmetric gradient domain ent algorithms, which are frequently applied to ML (sGDML)120,120 method or with smooth obtain a step-wise update of the weights: overlaps of atomic positions (SOAP)565 for GPR.119 wk+1 = wk − lr∇L2(w). (27) The gradient of the loss function as given in Neural Networks Another prominent ap- eq. (24) with respect to the weights is multiplied proach in ML is the use of NNs as highly flex- with a so-called learning rate, l . This hyper- ible parametric functions, which can fit huge r parameter is deemed one of the most important amounts of data and can map a molecular input hyperparameters used for training.9,566 In or- to many quantum chemical outputs.93 The sim- der to obtain an optimal solution, the learning plest form of NNs are multi-layer feed-forward rate needs to be chosen properly. Algorithms NNs, which are schematically represented in such as AdaGrad567 or Adam568 can automat- Fig. 10. As it is visible in Fig. 10, the width ically adapt the learning rate during training. Further, the second-order derivatives can be in- cluded into algorithms, which is for instance done in the global extended Kalman filter,569 in its parallel variant,570 or the element-decoupled variant.103 The loss function can be adapted so that more than only one property can be trained at once. This is often done to include the forces in the training process. In general, NNs possess various hyperparame- ters like the learning rate, regularizers, number of nodes, etc. As a consequence, an extensive Figure 10: Schematic representation of a multi- hyperparameter search complicates the use of layer feed-forward NN with inputs, X, nodes, n, NNs and makes them more complex to apply and outputs, Y . In the usual implementation than kernel methods. for the fitting of PESs, the NN maps a molec- Besides simple multi-layer feed-forward NNs, ular geometry to the ground state, which could high-dimensional variants exist. These net- be similarly done for any other single state. In works comprise several atomic NNs, which rep- case a manifold of excited states is described, resent atoms in their chemical and structural one molecular input can also be mapped to environment and are thus also called atomistic a vector of different excited states and addi- NNs. Each local atomic contribution, E , can tionally, other properties can be included. The a be summed up to provide the energy of the forces are treated as derivatives of the NN po- whole system, E, which is well known to work tentials with respect to Cartesian coordinates. for the ground state PESs: of the model is dependent on the number of NA t X nodes, nr, which are connected to each other E = Ea, (28) tu using weights, wrs. The indices refer to a con- a=1 nection between node r and node s from layer t and was originally implemented by Behler to and layer u, respectively. The number of nodes construct high-dimensional NN potentials.571 and hidden layers can be chosen independently Embedded-atom NNs533 are similar to high- of the training set size.

37 dimensional NNs in their way of constructing 5.2 Descriptors and Features the energy of a system. They differ in the Electronic structure methods can process and underlying descriptors to the ones of Behler. uniquely identify molecules using e.g. Carte- Atomic contributions to the energy are depen- sian coordinates. In contrast, such types of dent on the embedded density of atoms and are inputs are not optimal for ML models as the summed up according to eq 28. These embed- same molecular geometry, but translated or ro- ded density-like descriptors are approximated tated, could only be mapped to the same out- from atomic orbitals. put with great effort and unnecessary com- putational cost. Hence, a molecular descrip- Independent of a simple or an atomistic ar- tor should fulfill the following requirements: chitecture, the model can be used to fit a single It should be translationally, rotationally, and output or a vector of many outputs at the same permutationally invariant as well as differen- time. For ground state problems, a single-state tiable.102 It should also be unique with respect model is usually used, which maps an input to a to the relative spatial arrangement of atoms, single output, e.g. the PES of the ground state. universally applicable for any kind of system, Oftentimes, this single-state fashion is adapted and computationally efficient.552 However, a de- to fit different excited states with different NN scriptor can be more than that; it can already models.14,139,199,509 However, it has been shown include a part of the mapping, e.g., from a that including more excited-states in one model molecular structure to an energy. It can thus can be advantageous,93 as the excited-states are ease the task of the regressor and help to attain inherently linked to each other and so are the the best possible accuracy for a given training excited-state properties.37 Treating many ex- set. cited states can be referred to as multi-state The ways to represent a molecule to an ML model and the inclusion of more properties can model can be classified roughly into two cat- result in a multi-property model.76,93,95,186,400 egories: molecule-wise descriptors, which rep- The different properties can be weighted with resent the molecule as a whole to the ML respect to their magnitudes or importance for model, and atom-wise descriptors, which repre- a given chemical problem under investigation, sent atoms in their chemical and structural en- such that the best possible accuracy can be ob- vironment and build up a property using local tained.13 contributions.102,480 Both ways in describing a Another type of networks are convolutional molecular system have their merits and pitfalls NNs, which are most often applied in image and will be discussed along with their applica- or speech recognition,572–574 but can also be tions in recent studies for the excited states in adapted to process a molecular input and iden- the following. tify an optimal molecular descriptor. This type of network can be combined in an end-to-end fashion with an architecture, which fits this gen- Molecule-wise descriptors The distance erated molecular representation to a query out- matrix is one of the simplest descriptors that put.428,432,508,575,576 preserves rotational and translational invari- An important ingredient of all these ML mod- ance. Most often it is used in its inverse form els is the descriptor, which is mapped to the with distances between atoms a and b, output. In most studies, the descriptor is one 1 D = , (29) of many different possibilities to represent a ab || r − r || molecule, which will be discussed in the next a b section. giving rise to the symmetric inverse distance matrix, D. Due to the ill-definition of diagonal elements, which are not differentiable, the diag- onal elements are excluded and only the upper or lower triangular matrix is used to represent a

38 564 molecule to an ML model. Since the Hamil- another metric than the commonly used L1 tonian contains distances rather in the denom- or L2 norms can be employed, the so-called inator, it makes sense to also use the matrix of Wasserstein metric, which was tested with the inverse distances.92 The matrix of inverse dis- Coulomb matrix.581 tances is very similar to the Coulomb Matrix, Permutation invariant polynomials (PIPs), C:100 introduced by Bowman and co-workers,519,520,582 are frequently applied in a PIP-NN approach by ( 2.4 0.5Za if a = b Guo and coworkers to investigate photochemi- Cab = (30) ZaZb cal problems.141,142,145,359–362 The advantage of ||ra−rb|| these polynomials is that they are invariant to but the Coulomb matrix additionally considers permutation of atoms and inversion.145 They the atomic charges, Z. These types of descrip- comprise single-valued functions, pab, such as tors are frequently used in ML studies for the logarithmic or Morse like functions, which in- excited states. For example, MLMD simula- corporate internuclear distances, rab. The PIP tions in the excited states could be advanced us- vector, G is obtained applying a symmetriza- ing these simple descriptors92,93,137,138 and were tion operator, Sˆ, accounting for possible per- also accurate enough to fit NNs and KRR mod- mutation operations: els for excited-state properties.92,93,160,327,509,563 Distance based descriptors are further imple- NA ˆ Y mented in several program packages that have G = S pab (31) been used for photodynamics simulations with a

39 the FCHL representation.125,487 These represen- tems, such as DNA bases or amino acids. tations describe atoms in their chemical and structural local environment and usually rely Other types of descriptors Besides the on a cut-off function. This cut-off function benefits high-dimensional ML models offer for defines the sphere around an atom, which is the fitting of PESs of molecules, descriptors deemed to be important and is therefore con- are not restricted to the aforementioned exam- sidered when modelling the atomic local en- ples. In general, any type of descriptor might vironment. Radial distribution functions, so- be suitable for a given problem. Applied de- called second-order terms, account for inter- scriptors range from topological and binary atomic distances and are often used together features generated from SMILES strings584 to with angular distribution functions, i.e., third- normal modes, which are often used as a coor- order terms. It is further beneficial to in- dinate system and descriptors to fit diabatic clude first-order terms, i.e., the stoichiometry PESs.14,97,134,141,143,143–145,147,359–362,585 Other of atoms.125,428,446,583 Most often, higher order types of molecular features besides structure- terms than third-order terms are not included based ones, e.g. electronegativity, bond-order, due to increasing costs and little improvements oxidation states, ...,15,70 are also used. in accuracy.576 The description of PESs from atomic contri- Automatically generated descriptors butions is beneficial in order to treat systems of The selection of an optimal descriptor and the arbitrary sizes and to use systematic molecular optimization of the related parameters for this 107 fragmentation methods. Admittedly, the va- descriptor is no trivial task and requires ex- lidity of this approach is not so clear for the pert knowledge in many cases.576 A way to excited-states and consequently, such represen- circumvent an extensive parameter search is tations are less frequently used in ML studies offered by the aforementioned message pass- targeting the excited states. Up to day, only ing NNs,575 which include the descriptor pa- small molecules have been fitted with atom- rameters in the network architecture. In this wise representations, which are too small to way, they automatically fit the optimal param- prove the validity of excited-state PESs, which eters of a descriptor for a given problem, i.e., are constructed from local atomic contribu- training set under investigation. Such tailored tions. To the best of our knowledge, the largest descriptors can guarantee highly accurate so- molecule fitted with atom-wise descriptors con- lutions if the NN model is trained properly. 95 tained 12 atoms and was N -methylacetamide. PhysNet,586 HIP-NN587 or Deep Tensor NN + 13,93 139 Other molecules were CH2NH2 , CH2NH, (DTNN),508 which forms the basis of the deep 13 13 SO2 or CSH2. Further studies are needed learning model SchNet,,428,432 which in turn is to demonstrate whether an atom-wise construc- used within the SchNarc approach for excited tion of excited-state properties and PESs is pos- states,13 are examples of such NNs. sible or not. Nevertheless, this approach is most powerful for studies that aim to describe large and complex systems, which could poten- 6 Application of ML for Ex- tially be described from smaller building blocks. For instance, the construction of a DNA dou- cited States ble strand or a peptide could be, at least in In this chapter, we review ML studies of ex- principle, constructed from ML models that are cited states and their properties. We aim to trained on their smaller subsystems, i.e., DNA show how they have been employed to improve bases and amino acids, respectively. Unfortu- static and dynamics calculations and focus on nately, we are far away from having achieved a the used type of regressor, descriptor, train- description of large molecular systems for the ing set, and property. We will classify the ap- excited states, let alone the construction of ac- proaches according to Figure 1. curate PESs of medium-sized molecular sys-

40 6.1 Parameters for Quantum two ways, i.e., in a single-state fashion and in a Chemistry multi-state fashion.93 The applicability of such ML models to the simulation of photodynamics At the current state of research, the user must will be discussed. decide whether a multi-reference method is nec- essary or a single reference method is sufficient 6.3.1 ML in the Diabatic Basis to describe a chemical problem. It would be helpful if ML models could suggest a suitable Diabatic PESs are fitted with ML and related reference method, e.g. based on a literature methods since more than 25 years.149,377 An ad- search. Unfortunately, such a tool is not yet vantage of diabatic PESs is their smoothness, available, but ML can help to select an active which is perfectly matched by ML models built space for multi-reference methods. Jeong et. upon smooth functions. However, the tedious al70 developed an ML protocol for classification procedure to generate diabatic PESs remains. based on XGBoost558 to allow for a ”black box” Some effort is therefore devoted to develop ML- use of many multi-reference methods by auto- assisted diabatization procedures and eliminate matically selecting the relevant active space for this limiting step. molecular systems. The tedious selection of ac- tive orbitals and active electrons can thus be Diabatization Williams et. al140 incorpo- avoided. The accuracy of this approach was rated NNs into diabatization by ansatz and demonstrated for diatomic molecules in the dis- fit diabatic NO3 PESs. Recently, Shen and sociation limit and the molecules were repre- Yarkony94 fit two diabatic potentials of the cy- sented via the molecular orbital bond order and clopentoxy radical, C5H9O, and one state of − the average electronegativity of the system. cyclopentoxide, C5H9O , with 356 data points sampled from scans along different reaction co- 6.2 ML of Primary Outputs ordinates. The diabatization was assisted with NNs. Due to the high dimensionality of the To the best of our knowledge, no ML mod- system, the authors resort to application of els for providing primary outputs of quantum regularization in the fitting algorithm and an chemistry exist for excited states (see Figure 1). adapted loss function to obtain an accurate Targeting the primary output of a quantum representation of two-state diabatic PESs with chemistry simulation, i.e., the N-electron wave NNs. This novel strategy is envisioned for the function, or providing ML density (function- computation of the photoelectron spectrum of als) is far from trivial even for ground-state cyclopentoxide.94 Fitting 39 degrees of freedom problems.71–79,81,89,588–591 However, such an ap- in the diabatic basis is a huge improvement in proach for excited states could solve many prob- this research field. The authors further note lems and allow for wave function analysis, pro- that a comprehensive sampling of the full rel- viding additional insights like the excited state evant PESs in such high dimensional space is characters.592 Therefore, we expect such models problematic. to appear in the near future. Due to the aforementioned problems, a de- scription of medium-sized to large molecules 6.3 ML of Secondary Outputs with diabatic potentials is often done with more crude approximations.140,376 An exam- In the following, we summarize the contribu- ple is the LVC model,228 with its one-shot tions of ML models that fit the secondary variant,229 or the exciton model.177,593 For output of quantum chemical calculations, i.e., more details on this topic, the reader is re- PESs, SOCs, NACs, and transition as well as ferred to refs.63,228,313,594–596 The Frenkel exci- permanent dipole moments in the adiabatic and ton Hamiltonian can be used to describe light- diabatic basis (Figure 1). The prediction of the harvesting systems or charge-transfer.177,593 manifold quantities (see Fig. 2) can be done in Such a Hamiltonian was constructed for the

41 investigation of the excited state energies of authors note that the inclusion of more adia- bacteriochlorophylls of the Fenna-Matthews- batic states for the diabatization procedure and Olson complex. Multi-layer feed-forward NNs the consideration of additional relevant modes with the Coulomb matrix as a molecular de- can lead to more accurate results. All of the scriptor could accelerate the construction of reference simulations were carried out at the such Hamiltonians for the prediction of excited- CASSCF level of theory with KRR fitted di- state energies.160 The effective Hamiltonian of abatic PESs. the whole complex was subsequently used to In addition to KRR models, NNs were also predict excitation energy transfer times and ef- used to describe diabatic PESs. Seminal works ficiencies. Therefore, Häse et al. used exciton include PIP-based NNs by Guo, Yarkony and Hamiltonians as an input.159 co-workers. Absorption spectra and the dy- namics of excited states of NH3 and H2O could Fitting diabatic potentials and properties be studied by fitting potential energy matrix Given diabatic PESs, ML models can be used to elements.141,145,359–362,543 Subsequently, some of fit them. KRR models are often employed for the authors fit the dipole moments correspond- 1 14 this task, due to their ease of use and ability ing to the diabatic 1,2 A surface of NH3. to provide accurate predictions, as mentioned SOCs of formaldehyde were learned with NNs above. Recent studies by Habershon and co- in the diabatic picture.90 341 data points were workers focus on interpolation of diabatic PESs used for training of SOCs. A singlet and a and their use for grid-based quantum dynamics triplet state in the adiabatic basis were trans- methods, i.e., variational Gaussian wavepack- formed to diabatic states using Boys localiza- ets and MCTDH. The butatriene cation has tion.351 Since this diabatization is based on been investigated in two-dimensions comprising transition dipole moments, the respective prop- two electronic states.147 The description of this erties of the excited states had to be phase cor- molecule has been recently advanced with a new rected. The authors proved the accuracy of diabatization scheme, namely Procrustes diaba- their fitted PESs and emphasized the usability tization. The method was evaluated with two- of the ML models to describe full-dimensional state direct-dynamics MCTDH (DD-MCTDH) quantum dynamics.14,90,543 Very recently, they simulations of LiF and applied to four elec- investigated the OH + H2 reaction, i.e., the tronic states of butatriene.239 Some of the au- nonadiabatic quenching of the hydroxyl radical thors also carried out DD-MCTDH 4-mode/2- colliding with molecular hydrogen. Four dia- state143 and subsequently 12-mode/2-state dy- batic potentials including forces and couplings namics of pyrazine.144 The investigation of the were fitted using a least squares fitting proce- higher-dimensional space of pyrazine could be dure. 1345 data points of 1,2,3 2A adiabatic achieved by systematic tensor decomposition PESs were computed with MR-CISD.543 of KRR and advances conventional MCTDH The aforementioned ML models are single- simulations considerably with respect to accu- state models. Each energetic state and each racy and computational efficiency. Further, the coupling or dipole moment value resulting from method was applied to investigate the ultra- different pairs of states is fitted with a sepa- fast photodynamics of mycosporine-like amino rate ML model. While this yields justifiable acids, which are suitable as ingredients in sun- accuracy for energies and diabatic coupling val- screens due to their photochemical properties ues,93 dipole moments are vectorial properties and photostability.597 However, the reduced 6- and need to preserve rotational covariance.95 dimensional and 14-dimensional DD-MCTDH simulations with KRR interpolated PESs were As the aforementioned studies show, ML unable to reproduce the expected ultrafast pho- models are generally powerful to advance quan- todynamics, which had been observed in pre- tum dynamics simulations for the excited states viously performed surface hopping calculations and can also assist the construction of effec- and is typical for sunscreen ingredients. The tive Hamiltonians. However currently, diabatic

42 PESs cannot simply be fit for systems with arbi- a large amount of training data was required (> trary size and arbitrary complexity. The diaba- 65k data points). Coupling values were not fit- tization remains a methodological bottleneck, ted but, instead, the Zhu-Nakamura approach where additional developments are needed. was used to compute hopping probabilities. The investigation of medium-sized to larger Later, Dral et al.138 applied KRR models to molecular systems, especially the investigation accurately fit a two-state spin-Boson Hamilto- of their temporal evolution, is more often car- nian and reproduce reference dynamics using ried out in the adiabatic basis using on-the- 1,000 and 10,000 data points. NAC vectors fly simulations. An increasing number of re- were fit in a single-state fashion. During dy- cent studies focus on fitting such adiabatic namics simulations, conformations close to crit- PESs. The inconsistencies in adiabatic proper- ical regions were computed with the reference ties make such quantities generally more chal- method instead of the ML model in order to lenging to fit, which is why this field of research allow for accurate transitions. gained a lot of attention relatively late, i.e., only In another study, Chen et al.139 used two sep- in the last 3 years. arate deep NNs to fit the energies and forces of two adiabatic singlet states of CH2NH. About 6.3.2 ML in the Adiabatic Basis 90k data points were used to generate these single-state models. Using the Zhu-Nakamura Surface hopping MD Probably, the first approach to account for hopping probabilities, ML models for MQCD calculations date back to the reference dynamics could be reproduced 96 the year 2008. Nonadiabatic MD simulations and quantum chemical calculations were re- were carried out with NN-interpolated PESs placed completely during the dynamics. to investigate O2 scattered from Al(III). Sym- Cui and coworkers601 further developed a 598 metry functions were used as descriptors. A multi-layer energy-based fragmentation method spin-unpolarized singlet and a spin-polarized to study the excited-state dynamics and pho- triplet state at DFT level of theory were fit- tochemistry of larger systems. This scheme 598,599 ted with 3768 data points. This two-state composes a molecular system into a photo- spin-diabatic problem allowed for evaluation of chemically active (inner) region and a photo- coupling values and singlet-triplet transitions chemically inert (outer) region. In the original with the fewest switches surface hopping ap- scheme, the active region and the interactions 403,404 proach. In a later study, another adia- with the outer region are described with the batic spin-polarized PES was included and cou- multi-reference method CASSCF, whereas the pling values were computed between singlets outer region is treated with DFT. This decom- 600 and triplets and evaluated from constructed position of the total energy of a system allows to 91 Hamiltonian matrices. MD simulations were treat larger systems, which cannot be described executed using a manifold of ML-fitted PESs fully with CASSCF. The approach is similar to according to different spin-configurations. The QM/MM (quantum mechanics/molecular me- studies showed that singlet-triplet transitions chanics) schemes in the mechanical embedding are highly probable during the scattering event framework. The authors simulated two-state 91,96 of O2 on Au(III). photodynamics of CH3N=NCH3 (inner region) After these two seminal studies, the interest including five water molecules (outer region) in advancing MQC photodynamics simulations without the use of ML. The Zhu-Nakamura ap- in the adiabatic basis increased mainly in the proximation to model hopping probabilities in last three years. One of the first works dur- nonadiabatic MD simulations was applied.601 137 ing this time was conducted by Hu et. al, In order to make the simulations more effi- who investigated the nonadiabatic dynamics of cient, the authors replaced the DFT calcula- 6-aminopyrimidine with KRR and the Coulomb tions with deep multi-layer feed-forward NNs matrix. Due to the many degrees of freedom of using a distance-based descriptor,123 hence they the molecule and including three singlet states, describe the ground state energies and forces

43 of the photochemically inert region with ML KRR models was proposed to be a result of the and describe the S1 and S0 state of the inner parametric dependence of the depth of NNs and region with CASSCF. The hybrid ML multi- the non-parametric dependence of the depth of layer energy-based fragmentation method can KRR models. Results further suggested that reproduce the photodynamics of the system.443 small differences between the reference method Subsequently, the deep NNs were replaced with and ML models, especially in critical regions of embedded-atom NNs533 and accurate second the PESs, can lead to completely wrong pho- derivatives could be computed efficiently.444 todynamics simulations.93 Nevertheless, multi- Recently, we sought to fit NACs and transi- reference quantum chemical potential energy tion and permanent dipole moments in addition curves could be faithfully reproduced with KRR to energies and forces of three singlet states of models and NN models for the three singlet en- + + the methylenimmonium cation, CH2NH2 , using ergies of CH2NH2 . deep NNs and the matrix of inverse distances In order to omit the extensive hyperpa- as a molecular descriptor.92 We were able to rameter search of the descriptor and regres- perform ML-enhanced excited-state MD simu- sor, we further developed the SchNarc ap- lations with hopping probabilities based on ML- proach for photodynamics,13 which is based fitted NACs. NNs could replace the reference on SchNet.428,432 SchNarc allows for (1) a de- method MR-CISD completely during the dy- scription of SOCs, (2) an NAC approximation namics. Long time scale photodynamics simu- based on ML-fitted PESs, their first and sec- lations for 1 ns were achieved using the mean ond derivatives with respect to Cartesian coor- of 2 NN models in approximately two months, dinates, and (3) a phase-free training algorithm whereas the reference method would have taken to enable a training of raw quantum chemical an estimated 19 years to compute the dynam- data. The SchNarc approach is based on the ics for 1 ns on the same computer. This study message passing NN SchNet,428,432 which was demonstrated the possibility of MLMD simula- adapted by us for the treatment of a manifold tions to go beyond time scales of conventional of excited electronic states. Additionally, this methods. As another benefit of the ML models, model can describe dipole moments using the it was shown that a large ensemble of trajecto- charge model of ref,109 also adapted for excited- ries could be calculated, still at lower cost than states. All excited-state properties can be de- a few trajectories with the reference method.92 scribed in one ML model in a multi-state fash- With the same training set, we further as- ion. The performance of SchNarc was evaluated sessed the performance of KRR together with with surface hopping dynamics: Three singlet 93 von Lilienfeld and co-workers. The opera- and three triplet states of SO2 were computed tor formalism602 and the FCHL representa- with ML models for 700 fs and the underlying tion125,487 were used to fit the three singlet PESs were based on an "one-shot" LVC(MR- + 229 states of CH2NH2 . A single-state treatment CISD) model. CSH2 was investigated using 2 and a multi-state treatment for predicting en- singlets and 2 triplet states for 3 ps at CASSCF ergies were compared. To this aim, a multi- level of theory representing slow population state KRR approach as developed with an addi- transfer, and the performance of SchNarc to tional kernel that encodes the quantum energy reproduce ultrafast transitions during dynam- + levels. The accuracy of KRR models could be ics was assessed using CH2NH2 with the afore- improved using this extended approach.93 The mentioned training set. The hopping proba- KRR models were further compared to deep bilities were computed according to ML-fitted NN models regarding their ability to predict SOCs and NACs – the latter being fitted in a dipole moments and NACs. While NNs yielded rotationally covariant way as derivatives of vir- slightly higher accuracy at the largest avail- tual ML properties and approximated from ML able training set size, KRR models exhibited PESs. In all cases, excellent agreement with a steeper learning curve, hence more efficient the reference method could be achieved. No- learning. The different performance of NNs and ticeably, all the aforementioned photodynam-

44 ics studies with ML models13,92,93,137–139 make by slow population transfer. Hence, less Hes- use of Tully’s fewest switches surface hopping sian evaluations are required to estimate the approach with hopping probabilities based on hopping probabilities. coupling values or approximated schemes.403,404 The time required to train a SchNarc model on a GeForce GTX 1080 Ti GPU is approxi- Exemplary timings for MLMD, LVC dy- mately 11 hours for energies and forces of 3 sin- + namics, and MQCD The speed-up of simu- glet states with 3,000 data points of CH2NH2 , lations is one of the main arguments employed about 13 hours for energies, forces, and SOCs for promoting ML in quantum chemistry. In or- of 2 singlet and 2 triplet states using 4,000 data der to get an idea about the computational time points of CSH2 and about 4 hours for energies used in different calculations, we provide an ex- and forces of 3 singlet states of SO2 using 5,000 ample here. The timings of surface hopping data points. MD with analytical PESs (from LVC), quantum chemical PESs, and ML-fitted PESs based on Table 2: Comparison of the timings to compute fitted and approximated NACs from Hessians 100 fs with the surface hopping including arbi- trary couplings (SHARC)242,344,466 method. For can be found for three exemplary molecules in + Table 2. SO2 and CH2NH2 , three singlet states are de- Obviously, crude excited-state force fields like scribed and for CSH2 two singlet and two triplet the LVC model are faster than ML models, e.g., states. The molecule SO2 is approximated us- ing a highly efficient LVC model,229 while the for SO2. We note that even such force field underlying reference method to describe the implementations can probably still be stream- + lined for speed but will always be more expen- excited states of CH2NH2 is MR-CISD/aug- sive than ground-state MD simulations, where cc-pVDZ and of CSH2 is CASSCF(6,5)/def2- it would take approximately 0.005 seconds to SVP. SchNarc is used for the MLMD sim- simulate 100 fs for the gas-phase methylenim- ulations. Once, energies, forces, and NACs + are trained and predicted (MLMD1) and once, monium cation, CH2NH2 , using a state-of-the- art program like Amber.208 NACs are approximated from first- and second- However, dynamics based on highly accurate order derivatives of ML PESs (MLMD2). 2x 13 quantum chemical calculations can be acceler- Intel Xeon E5-2650 v3 CPUs are used. ated significantly with ML-fitted PESs, e.g., + 100 fs dynamics [s/CPU] SchNarc models for CH2NH2 based on MR- MLMD1 MLMD2 Reference CISD/aug-cc-pVDZ.13 The speedup is higher if SO 10 12 2-3 NACs are learned directly (MLMD1) compared 2 CH NH+ 24 250 74,224 to when they are approximated from Hessians 2 2 CSH 14 16 104 (MLMD2). A lot of Hessian evaluations are re- 2 quired in this example because ultrafast tran- + sitions occur in CH2NH2 . The second-order derivatives reduce the efficiency by a factor of Dipole Moments In addition to the inves- about ten. Nevertheless, Hessian calculations tigation of the temporal evolution of some sys- of ML-PESs can be accelerated by a factor of tems in the excited states, permanent and tran- about 5-10 using a GPU (dependent on the sition dipole moments have been computed molecule and GPU used). with ML models. As mentioned before, in our Table 2 further shows that a cheaper underly- earlier approaches, we fitted permanent and ing reference method, such as CASSCF(6,5)/def2- transition dipole moments as single values with NNs and KRR – strictly speaking we were ne- SVP used for CSH2, does not allow for such a significant speed-up. In this example however, glecting the rotational covariance of the vec- the difference between simulations with learned tors (since rotations were negligible in these 92,93 NACs and approximated NACs is small be- simulations). The SchNarc model improved on this description by treating dipole moments cause the dynamics of CSH2 is characterized

45 as vectorial properties. The NN and KRR 6.4 ML of Tertiary Outputs models for dipole moments have been evalu- The secondary outputs, such as dipole moments ated and compared to quantum chemical refer- or excited state energies can be used to cal- ence dipole moments using learning curves and culate oscillator strengths (eq 1) and energy MAEs. Their potential to compute UV spectra gaps (Fig. 1(d)). These properties can serve for was emphasized. the modelling of UV absorption spectra. UV The use of dipole moments to actually spectra were computed in the previously de- simulate UV spectra was demonstrated by scribed studies of N -methylacetamid with the Jiang, Mukamel, and co-workers using N - ML fitted transition dipole moments. Jiang, methylacetamide, a model system to investi- Mukamel and co-workers509 applied the tran- gate peptide bonds.95,509 They evaluated the sition dipole moment and additionally fitted ability of ML to describe transition dipole mo- nπ∗ and ππ∗ excitation energies to compute ments at TDDFT level of theory. In a first UV spectra this molecule with NNs. Subse- attempt,509 the authors predicted dipole vec- quently, some of the authors95 used these ex- tors as independent values. 14 internal coor- citation energies and the transition dipole mo- dinates in combination with multi-layer feed- ments to model a Frenkel exciton Hamiltonian forward NNs were used to predict transition for proteins using amino acid residues and pep- energies of N -methylacetamide. Xyz represen- tide bonds. This effective Hamiltonian could tations served as an input for fitting ground subsequently be used to approximate UV spec- state dipole moments. The Coulomb matrix tra of proteins. The interaction between amino was employed to fit transition dipole moments acid residues and peptides was neglected so only for the nπ∗ and ππ∗ transitions, but did not the isolated peptide excitation energies, i.e., lead to sufficiently accurate results. Higher ac- those of N -methylacetamid, and the respective curacy was obtained by replacing the atomic transition dipole moments were needed to con- charges in the Coulomb matrix (eq 30) with struct the Hamiltonian. The authors made use charges from natural population analysis. The of the dipole-dipole approximation603 and ap- choice of descriptors was justified by screen- plied embedded-atom NNs. ing different types of descriptors for prediction Ramakrishnan et. al327 predicted excitation of different properties. In a later work, some energies of the lowest-lying two excited singlet of the authors used embedded-atom NNs to states, S and S , as well as corresponding os- predict transition dipole moments from atomic 1 2 cillator strengths obtained from TDDFT calcu- contributions in a rotationally covariant way. lations with KRR. The QM8496 data base was The dipole moment vector between two states used consisting of 20k organic molecules. With i and j was obtained as a linear combination of the ∆-learning approach, CC2 accuracy could three contributions: be obtained. Very recently, Xue et al.563 as- i j 3 sessed the performance of KRR models with µij = µT + µ + µT (33) T the normalized inverse distances as a molecu- i j lar descriptor to predict absorption spectra of µT and µT were modeled using the charge model 3 benzene and a derivative of acridine contain- of ref 109. A third contribution, µT , was ob- i j ing 38 atoms. Therefore, the authors learned tained as the cross product of µT and µT : the excited-state energy gaps of several states NA and the corresponding oscillator strengths in a 3 X 3 i j µT = qa(µT × µT ) (34) single-state fashion. Applying Gaussian broad- a ening, the absorption cross sections could be j computed at TDDFT accuracy. µi , µ and q3 were outputs of the same T T a Pronobis et al.156 compared 2-body, 3-body embedded-atom NN. and automatically designed descriptors to learn TDDFT HOMO-LUMO gaps as well as first

46 and second vertical excitation energies. More ening. Geometries from the QM7b494,495 and than 20k molecules of the QM9 data base424,496 QM9424,496 data base were used for training were selected for this purpose and learning and molecular spectra were tested using 10k curves were used to evaluate the learning be- additional diastereomers, which were also used haviour of different ML models. While atom- by Ramakrishnan et. al327 to evaluate the wise descriptors worked well for HOMO-LUMO ∆-learning approach. The convolutional NNs gaps, the authors concluded that the accu- with the Coulomb matrix and DTNNs with racy of predicted transition energies is not an automatically generated representation out- sufficiently accurate and suggested that ad- performed the simpler NNs. Overall, good vanced non-local descriptors might be neces- agreement to reference DFT spectra could be sary to achieve higher accuracy. They fur- achieved.150 ther proposed the idea of encoding information Markland and co-workers447 trained NNs about the electronic state in the ML model.156 with atom-centered Chebyshev polynomial de- Indeed, our recent study, in which we com- scriptors108 on the TDDFT/CAM-B3LYP/6- pared the performance of KRR and NN models 31+G* S0-S1 energy gap of the deprotonated with atom-wise and molecule-wise descriptors trans-thiophenyl-p-coumarate (chromophore of demonstrated that encoding of the energy level yellow protein) in water and Nile red chro- is advantageous.93 mophore in water and benzene. Farthest point Recently, Kang et. al584 used 500,000 sampling121 was used to select about 2,000 data molecules of the PubChemQC502 data base to points from a larger set of 36,000 data points train a random forest model on the excita- and was compared to random sampling. The tion energy and the oscillator strength corre- authors assessed the performance of three dif- sponding to the electronic state with the high- ferent ML approaches to compute absorption est oscillator strength. 10 singlet states, as spectra, spectral densities and 2-dimensional available in the PubChemQC data base, were electronic spectra. One model (hidden solva- evaluated for that purpose. The authors used tion) completely ignored any environmental ef- SMILES (simplified molecular-input line-entry fects and only described the chromophore, an- system) strings and converted them into de- other model (indirect solvation) incorporated scriptors. The descriptors comprised several environmental effects within a 5Å cutoff of the topological604 and binary605 fingerprints, which atomistic descriptor for the chromophore and were calculated with the help of the RDkit a third model (direct solvation) treated the library.606 The authors compared the predic- whole system, i.e., the chromophore and the tion accuracy to the aforementioned models and atoms of the solvent, explicitly. As expected, stated that their model outperformed previous the hidden solvation model turned out to be ML models in the task of predicting accurate os- insufficiently accurate for systems with strong cillator strengths and excitation energies for the solvent-chromophore interactions, but was com- most probable transition in organic molecules. parable to the hidden solvation model when de- Analysis of important features led the authors scribing Nile red chromophore in benzene. The identify that nitrogen-containing heterocycles indirect solvation and direct solvation models are important for high oscillator strengths in were comparable to each other, but with respect molecules. The authors concluded that their to the computational efficiency, the indirect sol- study could serve the design of new fluorophores vation model was beneficial. This model could with high oscillator strengths.584 reproduce reference linear absorption spectra, Ghosh et. al150 used multi-layer feed-forward spectral densities, and could capture spectral NNs, convolutional NNs and DTNNs to fit 16 diffusion of 2-dimensional electronic spectra of highest occupied orbital energies from DFT, all treated chromophores.447 i.e., the respective eigenvalues, for the compu- Penfold and co-workers153 applied deep multi- tation of molecular spectra with a full width at layer feed-forward NNs to proof the ability of half maximum of 0.5 eV for Gaussian broad- ML to predict X-ray absorption spectra (XAS),

47 which provide a wealth of information on the to evaluate catalytic and material properties of geometry and electronic structure of chemical metal complexes. Descriptors based on a selec- systems, especially in the near-edge structure tion of empirical features were used to capture region. Note that X-Ray free-electron laser the bonding in inorganic molecular systems. spectroscopy can further be used to generate The performance of descriptors including dif- ultrashort X-ray pulses to investigate photo- ferent features was assessed for a set of octahe- dynamics simulations in real-time. The train- dral complexes with first-row transition metals. ing set for the prediction of Fe K-edge X- The most important features were identified to ray near-edge structure spectra contained 9040 be the atom, which connects the ligand to the data points. The inputs for NNs were gener- metal, its environment and its electronegativ- ated using local radial distributions around the ity, the metal identity and its oxidation state, Fe absorption site of arbitrary systems taken as well as the formal charge and denticity of from the Materials Project Database.607 Qual- the ligand.612 The ML models were tested on itatively accurate peak positions and intensi- spin-crossover complexes and could assign the ties could be obtained computationally efficient correct spin in most cases. Additionally, ML and the structural refinement of nitrosylmyo- models were applied for the discovery of inor- 2+ 613–616 globin and [Fe(bpy)3] was assessed with NNs. ganic complexes The authors noted that future development The inverse design of molecules with specific is needed to accurately capture structures far properties was further targeted by Schütt et. from equilibrium as well as irregularities in the al,76 who developed SchNOrb, a deep NN model bulk. based on SchNet. The automatically generated Another study was executed by Aarva et descriptor was extended with a description of al.,608 who focused on XAS and X-ray pho- atom pairs in their chemical and structural en- toelectron spectra of functionalized amorphous vironment. An analytic representation of the carbonaceous materials. By clustering of DFT electronic structure of a molecular system was data with unsupervised ML techniques aver- obtained in a local atomic orbital representa- age fingerprint spectra of distinct functional- tion. The analytic derivatives of the electronic ized surfaces could be obtained. The authors structure allowed for optimization of electronic use GPR. Similarly to the aforementioned state properties. This was demonstrated by minimiz- encoding,93 the authors encoded the electronic ing and maximizing the HOMO-LUMO gap of structure, i.e., the ∆-Kohn Sham values (core- malonaldehyde.486 Besides, the ML method was electron binding energies), in a Gaussian kernel. used to predict the lowest 20 molecular orbitals This kernel was then linearly combined with a of ethanol at DFT level of theory, to investigate structure-based kernel based on the SOAP609 proton transfer in malonaldehyde using ground- descriptor. The spectra computed from the state dynamics and to analyze bond order and different clusters were used to fit experimental partial charges of uracil. spectra allowing for an approximation to the Bayesian NN models were applied by Häse et. composition of experimental samples on a semi- al158 to relate molecular geometries to the out- quantitative level. The so-called fingerprint come of nonadiabatic MD simulations obtained spectra, which enabled the differentiation of the with CASSCF. Normal modes with and with- spectral signatures, were assessed in a previous out velocities of initial conditions served as an study using different models for amorphous car- input for NN models. Velocities in addition to bon,610 among them an ML fitted PES using normal modes as descriptors improved the ac- GPR.110,611 curacy of ML models slightly, pointing out that Kulik and co-workers15 used deep NNs to pre- normal modes contain already enough informa- dict the spin-state ordering in transition metal tion for the sake of their study. The dissociation complexes to determine the spin of the low- times of 1,2-dioxetane obtained from nonadia- est lying energetic state in open-shell systems. batic MD simulations was the targeted output. The determination of spin states is important The NNs could faithfully reproduce dissociation

48 times and further provided a measure of uncer- spectroscopy.619 The authors highlighted the tainty. The authors noted that their method applicability of their method to enhance stud- could be particularly interesting for analysis of ies on the optimization and design of opti- MLMD simulations. cal devices and further noted that their ap- proach can also be used to analyze transient 6.5 ML-Assisted Analysis absorption spectra. Aspuru-Guzik and co- workers152 applied Bayesian NNs to find corre- The aforementioned studies have shown that lations of nanoaggregates with electronic cou- ML enables the simulation of MD simulations pling in semiconducting materials using ab- and spectra predictions at low computational sorption spectra. In general, the analysis of costs. The computational efficiency allows for experimental spectra and the inverse design of enhanced statistics, i.e., in case of MD simu- compounds is most frequently applied in the lations a huge number of trajectories and the research field of material science. Their de- simulations on long time scales.13,92 Therefore, scription goes beyond the scope of this review subsequent analyses of production runs can be- and the reader is referred to Refs 163–167,169. come a time limiting step of studies. This prob- lem was identified in the aforementioned study on the dissociation times of 1,2-dioxetane by 7 Conclusion and Future Häse et. al.158 Therefore, the authors further Perspectives used their method to interpret the outcomes of nonadiabatic MD simulations. 1,2-dioxetane In the last few years, machine learning (ML) is the target of their study as it is the small- has started to slowly enter the research field of est molecule known to show chemilumiescence photochemistry, especially the photochemistry after nonadiabatic transitions from an excited of molecular systems. Although this field of state to the ground state. The chemilumines- research is rather young compared to ML for cent properties of this compound were related the electronic ground-state, some groundbreak- to its decomposition rate into two formaldehyde ing works have already shown the potential of molecules. By analysis of the ML models that ML models to significantly accelerate and im- fit the dissociation times, correlations could be prove existing simulation techniques. So far, observed between the normal modes and the most studies provide a proof of concept using dissociation times. For example, the modes small molecular systems or model systems. Dif- corresponding to C-C bond stretching and C-O ferent applications are targeted and will also be bond stretching were relevant for the accurate aimed at in the future, ranging from dynamics prediction of dissociation times. It was further with excited-state ML potentials via absorption emphasized by the authors that although the spectra to the interpretation of data, see Fig. 1. findings of NNs were expected and obey phys- Analysing the different studies reviewed here, ical laws, ML models were helpful to extract some trends in the choice of reference methods, relevant information of large amount of data ML models, and descriptors can be observed. and could potentially serve as an inspiration to These trends are illustrated in Figure 11. humans. The pie chart in panel 11(a) shows the Time-resolved experimental photolumines- used reference methods for the computa- cence spectra could be analyzed with the Lu- tion of a training set to describe the excited miML software developed by Ðorđević et. al,617 states or excited-state properties of molecules. who applied linear regression models to learn As can be seen, about half of the training from computer-generated photoluminescence sets are computed with multi-reference meth- data. The software was employed to pre- ods.12–14,70,90,92–94,137,139–145,147,158,239,359,360,362,597 dict decay rate distributions618 of perovskite The employed single-reference approaches are nanocrystals from data generated with fem- exclusively based on DFT.15,76,91,95,96,150,156,327,447,467,509,584,610 tosecond broadband fluorescence upconversion Analytical methods or experimental data are

49 in full dimensions has not yet been investi- gated.137,143,144 Especially the huge number of data points is concerning in this case, as larger molecules with more energetic states and a complex photochemistry could require many more data points. A meaningful training set generation, which can be achieved with ac- tive learning, adaptive sampling and structure- based sampling techniques, is thus essential for dynamics simulations.92,109,479,480 Clustering of molecular geometries obtained from dynamics simulations with a cheap method further is ben- eficial for selecting important reference geome- tries.137,138,481–483 Still, the high costs and the complexity of multi-reference methods to com- pute an ample training set for ML also hampers the application of ML models to fit the excited states of larger polyatomic systems, whose ac- curate photochemical description is often addi- tionally complicated by a high density of elec- tronic states. Single reference methods, such as time- dependent DFT, are advantageous with respect Figure 11: Pie diagrams summarizing the ref- to the computational costs of the training set, erence methods used for the training set gen- but suffer from qualitatively incorrect PESs eration, the chosen ML models and the type in some conformational regions of molecules, of descriptors for the description of the excited such as dissociative regions. In principle, these states with ML. conformational regions could be excluded from the training set and the remaining conforma- tional space could be interpolated using ML, also applied.138,152,159,160,617 but the training set would then remain incom- When restricting the analysis to studies tar- plete and so would the dynamics. Schemes like geting dynamics, the fraction that employs the ∆−learning approach327 or transfer learn- multi-reference methods even increases. About ing329 could be helpful in this regard. These 70% of all dynamics studies use multi-reference approaches might be useful to let ML models methods to compute the training data for ML learn from single-reference data and adjust their models. 15% of the studies use single-reference accuracy according to multi-reference methods. methods and an equally large portion apply The direct use of approximated methods, such model Hamiltonians or analytical potentials. as time-dependent DFT-based tight binding, This shows that most chemical problems for the is most likely not suitable for photodynamics investigation of the excited states of molecules on long time scales, because such approaches require multi-reference accuracy. might easily be quantitatively incorrect. Of Recent studies of ML-based photodynam- particular concern is then the accumulation ics simulations have shown that many thou- of quantitatively tiny errors in the underlying sands of data points are necessary to describe potentials toward wrong dynamics trends. At a few excited-state potentials of small molec- the current stage of research, it is not clear ular systems. To the best of our knowl- whether such approximate potentials can pro- edge, the dynamics in the excited states with vide qualitatively correct trends for reaction ML for molecules with more than 12 atoms dynamics.400

50 In addition to the aforementioned problems, lated to the ground-state equilibrium structure the training set generation is complicated by of a molecular system or to electronic ground the arbitrariness of the signs of coupling val- state calculations, e.g. the HOMO-LUMO ues and properties resulting from two different gaps.76,150,156 Due to the limited transferabil- electronic states.13,14,90,92,93,95 This arbitrari- ity of existing ML models to predict the excited ness has to be removed in order to make data state PESs and properties of different molecular learnable with conventional methods. Such a systems, an extrapolation throughout chemical correction scheme is termed phase correction compound space is hindered in many cases. and has been applied to correct coupling values In order to fully exploit the advantages that and dipole moments.14,90,92,95,454 An alternative ML models offer and to achieve the aforemen- phase correction training algorithm has been tioned goal of a transferable ML model for the shown to be beneficial with respect to the costs excited states, a highly versatile descriptor is re- of the training set generation and has enabled quired, which can describe atoms in their chem- the learning of raw quantum chemical data.13 ical and structural environment and enables an ML model to treat molecules of arbitrary size Figure 11(b) shows which ML models are and composition. It would be highly desirable, applied in the discussed studies. About two if an ML model could then describe the pho- thirds rely on NNs, whereby simple multi- tochemistry of large systems, which are too ex- layer feed-forward NNs are most often em- pensive to compute with precise multi-reference ployed. Several research fields were advanced methods, using only small building blocks, i.e., with NN-fitted functions: photodynamics sim- small enough ones to describe their electronic ulations,13,91–93,96,138,139,141,142,145,359,360,362 spec- structure accurately. For example, the excited tra predictions and analysis,95,150,153,447,509,563 states of proteins or DNA strands could poten- excited-state properties,13–15,90,93,95,509 diabati- tially be predicted from contributions of amino zation procedures,94,140 interpretation of re- acids or DNA bases, respectively, which is most action outcomes,158,617 and the prediction often done using effective model Hamiltonians of HOMO-LUMO gaps or gaps between en- up to date.56 A local description of the excited- ergetic states.76,150,156 KRR methods were state PESs and their properties derived from mainly applied to interpolate diabatic poten- the ML-fitted PESs, could further provide a tials143,144,147,239,597 and in studies focusing on way toward excited-state ML/MM simulations more than one molecular systems.327 In general, alike QM/MM (quantum mechanics/molecular only a few studies focused on extrapolation mechanics) techniques.400,443,601 Unfortunately, throughout chemical compound space in the it is not yet known whether the excited-state excited states. Yet only the energies, HOMO- PESs and properties can be constructed from LUMO gaps or spectra based on fitted oscillator atomic contributions or not.400 strengths could be predicted using a single ML In studies comparing different ML models, it model for different molecules.15,153,156,327 Deci- was even suggested that non-local descriptors sion trees were used to select an active space might be needed or that the electronic state has for diatomic molecules.70 to be encoded explicitly in the molecular rep- One drawback of recently developed ML mod- resentation to enable a transferable description els is that they are molecule-specific and thus of the excited states with ML.93,156 not universal. In part, this issue is related To conclude, the reviewed studies focus on al- to the used molecular descriptors. As can be most all aspects of excited-state quantum chem- seen in panel (c) in Figure 11, most stud- istry and improve them successfully: ML mod- ies apply descriptors that capture molecules els can help to choose a proper active space as a whole. The few studies, which describe for multi-reference methods, they predict sec- PESs and properties of molecular systems from ondary and tertiary outputs of quantum chem- atomic contributions, either treat small molec- ical calculations and help in the interpreta- ular systems13,93,95 or predict properties re- tion of theoretical studies. ML models push

51 the boundaries of computed time scales92 and velop methods, which could develop into a uni- are used to investigate and analyze the huge versal approximator, make ML models perfectly amount of data we produce every day in ex- suited to advance this research field. The pos- periments or with high-performance comput- sibility of deep ML models to process a huge ers.158,617 amount of data can even assist the interpreta- It should be emphasized once more that the tion and analysis158,617 of many photochemical recent studies show that the goal of ML is not studies and can help to explore unknown physi- to replace existing methods completely, but to cal relations and be a source of potential human provide a way to improve them. In fact, ML inspiration. models for the excited states at their current Acknowledgement This work was finan- stage are far from replacing existing quantum cially supported by the Austrian Science Fund, chemical methods, and they are also far from W 1232 (MolTag) and the uni:docs program of being routine. Without human intervention, the University of Vienna (J.W.). P. M. thanks ML cannot solve existing problems and much the University of Vienna for continuous sup- remains to be done to describe systems beyond port, also in the frame of the research platform single, isolated molecules. ViRAPID. We thank P. A. Sánchez-Murcia for To the best of our knowledge, what is still help in setting up the quick Amber simulation missing is the proof that ML can provide an ap- for MD timings. proximation to the multi-reference wave func- tion of a molecular system. Such an achieve- ment would be a great advancement in the re- References search field of photochemistry, as any property we wish to know could possibly be derived from (1) Këpuska, V.; Bohouta, G. Next- the ML wave function. An ML representation Generation of Virtual Personal Assis- of the electronic structure would further be ben- tants (Microsoft Cortana, Apple Siri, eficial to allow for an inverse design of molecules Amazon Alexa and Google Home). 2018 with specific properties, which has been shown IEEE 8th Annual Computing and Com- to be feasible for the ground state of a molec- munication Workshop and Conference ular system.76 The optimization of photochem- (CCWC). 2018; pp 99–103. ical properties with respect to molecular ge- (2) Hoy, M. B. Alexa, Siri, Cortana, and ometries would be useful for many exciting re- More: An Introduction to Voice Assis- search fields, e.g. photocatalysis,165 photosen- tants. Med. Ref. Serv. Q. 2018, 37, 81– sitive drug design620 or photovoltaics.609,621 88. The multi-faceted photochemistry offers a perfect playground for ML models. It may be (3) Silver, D. et al. Mastering the Game of important to highlight that, despite the neg- Go with Deep Neural Networks and Tree ative image ML has suffered in some research Search. Nature 2016, 529, 484–489. communities, it cannot be denied that it opens up many new ways and possibilities to improve (4) Bansak, K.; Ferwerda, J.; Hain- simulations and make studies feasible that were mueller, J.; Dillon, A.; Hangartner, D.; considered unattainable only a few years, if not Lawrence, D.; Weinstein, J. Improving only months ago.442 The computational effi- Refugee Integration through Data- ciency and high flexibility of deep learning mod- Driven Algorithmic Assignment. Science els can lead this research field toward simula- 2018, 359, 325–329. tions of long time and large length scales. The (5) Leung, M. K. K.; Delong, A.; Ali- possibilities ML models offer are far from be- panahi, B.; Frey, B. J. Machine Learn- ing being exhausted. Considering the enormous ing in Genomic Medicine: A Review of chemical space, estimated to consist of more Computational Problems and Data Sets. 60 622 than 10 molecules, and the desire to de- Proc. IEEE 2016, 104, 176–197.

52 (6) Shen, D.; Wu, G.; Suk, H.-I. Deep Learn- (15) Taylor, M. G.; Yang, T.; Lin, S.; ing in Medical Image Analysis. Annu. Nandy, A.; Janet, J. P.; Duan, C.; Ku- Rev. Biomed. Eng. 2017, 19, 221–248. lik, H. J. Seeing Is Believing: Experimen- tal Spin States from Machine Learning (7) Chen, C.; Seff, A.; Kornhauser, A.; Model Structure Predictions. J. Phys. Xiao, J. DeepDriving: Learning Af- Chem. A 2020, 124, 3286–3299. fordance for Direct Perception in Au- tonomous Driving. The IEEE Interna- (16) Kulik, H. J. Making Machine Learning a tional Conference on Computer Vision Useful Tool in the Accelerated Discovery (ICCV). 2015. of Transition Metal Complexes. WIREs Comput. Mol. Sci. 2020, 10, e1439. (8) Yang, X.; Wang, Y.; Byrne, R.; Schnei- der, G.; Yang, S. Concepts of Artificial (17) Power, P. P. Stable Two-Coordinate, Intelligence for Computer-Assisted Drug Open-Shell (d1âĂŞd9) Transition Metal Discovery. Chem. Rev. 2019, 119, 10520– Complexes. Chem. Rev. 2012, 112, 10594. 3482–3507.

(9) Goodfellow, I.; Bengio, Y.; Courville, A. (18) Bousseksou, A.; MolnÃąr, G.; Ma- Deep Learning; MIT Press, 2016. touzenko, G. Switching of Molecular Spin States in Inorganic Complexes by Tem- (10) Gómez-Bombarelli, R.; Aspuru- perature, Pressure, Magnetic Field and Guzik, A. In Handbook of Materials Light: Towards Molecular Devices. Eur. Modeling : Methods: Theory and Mod- J. Inorg. chem. 2004, 2004, 4353–4369. eling; Andreoni, W., Yip, S., Eds.; Springer International Publishing: (19) Li, H.; Feng, H.; Sun, W.; Fan, Q.; Cham, 2018; pp 1–24. King, R. B.; Schaefer, H. F. First-Row Transition Metals in Binuclear Cyclopen- (11) Agrawal, A.; Choudhary, A. Perspective: tadienylmetal Derivatives of Tetram- Materials Informatics and Big Data: Re- ethyleneethane: η3, η3 versus η4, η4 alization of the âĂIJFourth ParadigmâĂİ LigandâĂŞMetal Bonding Related to of Science in Materials Science. APL Spin State and MetalâĂŞMetal Bonds. Mat. 2016, 4, 053208. Organometallics 2014, 33, 3489–3499.

(12) Schwilk, M.; Tahchieva, D. N.; von (20) Matsika, S.; Krylov, A. I. Introduction: Lilienfeld, O. A. Large yet Bounded: Theoretical Modeling of Excited State Spin Gap Ranges in Carbenes. arXiv Processes. Chem. Rev. 2018, 118, 6925– 2020, 2004.10600 . 6926.

(13) Westermayr, J.; Gastegger, M.; Marque- (21) Cohen, B.; Crespo-Hernández, C. E.; tand, P. Combining SchNet and SHARC: Hare, P. M.; Kohler, B. Ultrafast Excited- The SchNarc Machine Learning Ap- State Dynamics in DNA and RNA Poly- proach for Excited-State Dynamics. J. mers; Elsevier: Amsterdam, 2004; Chap- Phys. Chem. Lett. 2020, 11, 3828–3834. ter Ultrafast Excited-State Dynamics in (14) Guan, Y.; Guo, H.; Yarkony, D. R. Ex- DNA and RNA Polymers, pp 463–470. tending the Representation of Multistate (22) Levine, B. G.; Martínez, T. J. Isomer- Coupled Potential Energy Surfaces to In- ization Through Conical Intersections. clude Properties Operators using Neu- Annu. Rev. Phys. Chem. 2007, 58, 613– 1 ral Networks: Application to the 1,2 A 634. States of Ammonia. J. Chem. Theory Comput. 2020, 16, 302–313.

53 (23) Turro, N. J.; Ramamurthy, V.; Sca- (32) Casanova, D. Theoretical Modeling of iano, J. C. Principles of Molecular Pho- Singlet Fission. Chem. Rev. 2018, 118, tochemistry: An Introduction. 2009. 7164–7207.

(24) Yarkony, D. R. Nonadiabatic Quantum (33) Hestand, N. J.; Spano, F. C. Expanded Chemistry - Past, Present, and Future. Theory of H- and J-Molecular Aggre- Chem. Rev. 2012, 112, 481–498. gates: The Effects of Vibronic cou- pling and Intermolecular Charge Trans- (25) Barbatti, M.; Borin, A. C.; Ullrich, S. fer. Chem. Rev. 2018, 118, 7069–7163. Photoinduced Phenomena in Nucleic Acids I ; Topics in Current Chemistry; (34) Penfold, T. J.; Gindensperger, E.; Springer Berlin Heidelberg, 2014; Vol. Daniel, C.; Marian, C. M. Spin-Vibronic 355; pp 1–32. Mechanism for Intersystem Crossing. Chem. Rev. 2018, 118, 6975–7025. (26) Ibele, L. M.; Nicolson, A.; Cur- chod, B. F. E. Excited-State Dynamics of (35) Vacher, M.; Fdez. Galván, I.; Ding, B.- molecules with classically driven trajec- W.; Schramm, S.; Berraud-Pache, R.; tories and Gaussians. Mol. Phys. 2020, Naumov, P.; Ferré, N.; Liu, Y.-J.; Nav- 118, e1665199. izet, I.; Roca-Sanjuán, D.; Baader, W. J.; Lindh, R. Chemi- and Bioluminescence (27) Nelson, T. R.; White, A. J.; Bjor- of Cyclic Peroxides. Chem. Rev. 2018, gaard, J. A.; Sifain, A. E.; Zhang, Y.; 118, 6927–6974. Nebgen, B.; Fernandez-Alberti, S.; (36) Crespo-Otero, R.; Barbatti, M. Recent Mozyrsky, D.; Roitberg, A. E.; Tre- Advances and Perspectives on Nonadia- tiak, S. Non-adiabatic Excited-State batic Mixed QuantumâĂŞClassical Dy- Molecular Dynamics: Theory and Ap- namics. Chem. Rev. 2018, 118, 7026– plications for Modeling Photophysics in 7068. Extended Molecular Materials. Chem. Rev. 2020, 120, 2215–2287. (37) González, L.; Lindh, R. Quantum Chem- istry and Dynamics of Excited States : (28) Mai, S.; GonzÃąlez, L. Molecular Photo- Methods and Applications; John Wiley chemistry: Recent Developments in The- and Sons Ltd, 2020. ory. Angew. Chem. Int. Ed. 2020, n/a. (38) Harris, D. C.; Bertolucci, M. D. Symme- (29) Lischka, H.; Nachtigallová, D.; try and Spectroscopy: an Introduction to Aquino, A. J. A.; Szalay, P. G.; Vibrational and Electronic Spectroscopy; Plasser, F.; Machado, F. B. C.; Bar- New York: Dover Publications, 1989. batti, M. Multireference Approaches for Excited States of Molecules. Chem. Rev. (39) Ng, C.-Y. Vacuum Ultraviolet Pho- 2018, 118, 7293–7361. toionization and Photodissociation of Molecules and Clusters; World Scientific, (30) Ghosh, S.; Verma, P.; Cramer, C. J.; 1991. Gagliardi, L.; Truhlar, D. G. Combin- ing Wave Function Methods with Density (40) Zewail, A. H. Femtochemistry: Ultrafast Functional Theory for Excited States. Dynamics of the Chemical Bond; World Chem. Rev. 2018, 118, 7249–7292. Scientific, 1994; pp 3–22.

(31) Norman, P.; Dreuw, A. Simulating (41) Brixner, T.; Pfeifer, T.; Gerber, G.; Wol- X-Ray Spectroscopies and Calculating lenhaupt, M.; Baumert, T. In Femtosec- Core-Excited States of Molecules. Chem. ond Laser Spectroscopy; Hannaford, P., Rev. 2018, 118, 7208–7248. Ed.; Springer-Verlag: New York, 2005; pp 225–266.

54 (42) Iqbal, A.; Stavros, V. G. Active Partici- (50) Ashfold, M. N. R.; Bain, M.; pation of 1πσ* States in the Photodisso- Hansen, C. S.; Ingle, R. A.; Karsili, T. ciation of Tyrosine and its Subunits. J. N. V.; Marchetti, B.; Murdock, D. Phys. Chem. Lett. 2010, 1, 2274–2278. Exploring the Dynamics of the Pho- toinduced Ring-Opening of Heterocyclic (43) Kowalewski, M.; Fingerhut, B. P.; Dorf- Molecules. J. Phys. Chem. Lett. 2017, man, K. E.; Bennett, K.; Mukamel, S. 8, 3440–3451. Simulating Coherent Multidimensional Spectroscopy of Nonadiabatic Molecu- (51) Tajti, A.; Fogarasi, G.; Szalay, P. G. lar Processes: From the Infrared to the Reinterpretation of the UV Spectrum of X-Ray Regime. Chem. Rev. 2017, 117, Cytosine: Only Two Electronic Transi- 12165–12226. tions? ChemPhysChem 2009, 10, 1603– 1606. (44) Soorkia, S.; Jouvet, C.; Grégoire, G. UV Photoinduced Dynamics of Conformer- (52) Barbatti, M.; Szymczak, J. J.; Resolved Aromatic Peptides. Chem. Rev. Aquino, A. J. A.; Nachtigallová, D.; 2020, 120, 3296âĂŞ3327. Lischka, H. The Decay Mechanism of Photoexcited Guanine – A Nonadiabatic (45) Yusong Liu, et al., Spectroscopic and Dynamics Study. J. Chem. Phys. 2011, Structural Probing of Excited-State 134, 014304. Molecular Dynamics with Time-Resolved Photoelectron Spectroscopy and Ultra- (53) Lu, Y.; Lan, Z.; Thiel, W. Photoinduced fast Electron Diffraction. Phys. Rev. X Phenomena in Nucleic Acids II ; Topics 2020, 10, 021016. in Current Chemistry; Springer Berlin Heidelberg, 2014; Vol. 356; pp 89–122. (46) Martínez, T. J. Insights for Light-Driven Molecular Devices from Ab Initio Mul- (54) Ruckenbauer, M.; Mai, S.; Marque- tiple Spawning Excited-State Dynamics tand, P.; González, L. Photoelectron of Organic and Biological Chromophores. Spectra of 2-Thiouracil, 4-Thiouracil, Acc. Chem. Res. 2006, 39, 119–126. and 2,4-Dithiouracil. J. Chem. Phys. 2016, 144, 074303. (47) Barbatti, M.; Sellner, B.; Aquino, A. J. A.; Lischka, H. In Radiation Induced (55) Manathunga, M.; Yang, X.; Luk, H. L.; Molecular Phenomena in Nucleic Acids; Gozem, S.; Frutos, L. M.; Valentini, A.; Shukla, M., Leszczynski, J., Eds.; Chal- Ferrè, N.; Olivucci, M. Probing the Pho- lenges and Advances in Computational todynamics of Rhodopsins with Reduced Chemistry and Physics; Springer Nether- Retinal Chromophores. J. Chem. Theory lands, 2008; Vol. 5; pp 209–235. Comput. 2016, 12, 839–850.

(48) Subotnik, J. E.; Jain, A.; Landry, B.; (56) Nogueira, J. J.; Plasser, F.; González, L. Petit, A.; Ouyang, W.; Bellonzi, N. Un- Electronic Delocalization, Charge Trans- derstanding the Surface Hopping View of fer and Hypochromism in the UV Ab- Electronic Transitions and Decoherence. sorption Spectrum of Polyadenine Un- Annu. Rev. Phys. Chem. 2016, 67, 387– ravelled by Multiscale Computations 417. and Quantitative Wavefunction Analysis. Chem. Sci. 2017, 8, 5682–5691. (49) Curchod, B. F. E.; Martínez, T. J. Ab Initio Nonadiabatic Quantum Molecular (57) Mai, S.; Mohamadzade, A.; Marque- Dynamics. Chem. Rev. 2018, 118, 3305– tand, P.; González, L.; Ullrich, S. Sim- 3336. ulated and Experimental Time-Resolved

55 Photoelectron Spectra of the Intersys- Dynamics of Azobenzene Photoisomer- tem Crossing Dynamics in 2-Thiouracil. ization. J. Am. Chem. Soc. 2003, 125, Molecules 2018, 23, 2836. 8098–8099.

(58) Rauer, C.; Nogueira, J. J.; Marque- (66) Toniolo, A.; Olsen, S.; Manohar, L.; tand, P.; González, L. Stepwise photosen- Martínez, T. J. Conical Intersection Dy- sitized thymine dimerization mediated namics in Solution: The Chromophore of by an exciton intermediate. Monatsh. Green Fluorescent Protein. Faraday Dis- Chem. 2018, 149, 1âĂŞ9. cuss. 2004, 127, 149–163.

(59) Zobel, J. P.; Heindl, M.; Nogueira, J. J.; (67) Domcke, W.; Yarkony, D.; Köppel, H. González, L. Vibrational Sampling and Conical Intersections: Theory, Compu- Solvent Effects on the Electronic Struc- tation and Experiment; Advanced Series ture of the Absorption Spectrum of in Physical Chemistry; World Scientific 2-Nitronaphthalene. J. Chem. Theory Publishing Company, 2011. Comput. 2018, 14, 3205–3217. (68) Serrano-Andrés, L.; Merchán, M. Quan- (60) Pathak, S. et al. Tracking the Ultravio- tum chemistry of the excited state: 2005 let Photochemistry of Thiophenone Dur- overview. J. Mol. Struc.-THEOCHEM ing and Beyond the Initial Ultrafast Ring 2005, 729, 99 – 108. Opening. 2019. (69) Chandrasekaran, A.; Kamal, D.; Ba- (61) Maria Teresa Neves-Petersen, S. P.; tra, R.; Kim, C.; Chen, L.; Ram- Gajula, G. P. UV Light Effects on prasad, R. Solving the Electronic Struc- Proteins: From Photochemistry to ture Problem with Machine Learning. npj Nanomedicine, Molecular Photochem- Comput. Mater. 2019, 5, 22. istry - Various Aspects; IntechOpen, 2012; Chapter 7. (70) Jeong, W.; Stoneburner, S. J.; King, D.; Li, R.; Walker, A.; Lindh, R.; (62) Cadet, J.; Grand, A.; Douki, T. Photoin- Gagliardi, L. Automation of Active duced Phenomena in Nucleic Acids II ; Space Selection for Multireference Meth- Topics in Current Chemistry; Springer ods via Machine Learning on Chemical Berlin Heidelberg, 2014; Vol. 356; pp Bond Dissociation. J. Chem. Theory 249–275. Comput. 2020, 16, 2389–2399.

(63) Segatta, F.; Cupellini, L.; Garavelli, M.; (71) Carleo, G.; Troyer, M. Solving the Quan- Mennucci, B. Quantum Chemical Model- tum Many-Body Problem with Artifi- ing of the Photoinduced Activity of Mul- cial Neural Networks. Science 2017, 355, tichromophoric Biosystems. Chem. Rev. 602–606. 2019, 119, 9361–9380. (72) Saito, H. Solving the BoseâĂŞHubbard (64) Landry, B. R.; Subotnik, J. E. Quan- Model with Machine Learning. J. Phys. tifying the Lifetime of Triplet Energy Soc. Jpn. 2017, 86, 093001. Transfer Processes in Organic Chro- mophores: A Case Study of 4-(2- (73) Nomura, Y.; Darmawan, A. S.; Ya- Naphthylmethyl)benzaldehyde. J. Chem. maji, Y.; Imada, M. Restricted Boltz- Theory Comput. 2014, 10, 4253–4263. mann Machine Learning for solving strongly correlated quantum systems. (65) Schultz, T.; Quenneville, J.; Levine, B.; Phys. Rev. B 2017, 96, 205152. Toniolo, A.; Martínez, T. J.; Lochbrun- ner, S.; Schmitt, M.; Shaffer, J. P.; Zgier- ski, M. Z.; Stolow, A. Mechanism and

56 (74) Han, J.; Zhang, L.; E, W. Solving (83) Nelson, J.; Tiwari, R.; Sanvito, S. Ma- Many-Electron Schrödinger Equation us- chine Learning Density Functional The- ing Deep Neural Networks. J. Comput. ory for the Hubbard Model. Phys. Rev. Phys. 2019, 399, 108929. B 2019, 99, 075132.

(75) Townsend, J.; Vogiatzis, K. D. Data- (84) Cheng, L.; Welborn, M.; Chris- Driven Acceleration of the coupled- tensen, A. S.; Miller, T. F. A Uni- Cluster Singles and Doubles Iterative versal Density Matrix Functional from Solver. J. Phys. Chem. Lett. 2019, 10, Molecular Orbital-Based Machine 4129–4135. Learning: Transferability Across Or- ganic Molecules. J. Chem. Phys. 2019, (76) Schütt, K. T.; Gastegger, M.; 150, 131103. Tkatchenko, A.; Müller, K.-R.; Mau- rer, R. J. Unifying Machine Learning and (85) Lei, X.; Medford, A. J. Design and quantum chemistry with a deep neural Analysis of Machine Learning Exchange- network for molecular wavefunctions. Correlation Functionals via Rotation- Nat. Commun. 2019, 10, 5024. ally Invariant Convolutional Descriptors. Phys. Rev. Materials 2019, 3, 063801. (77) Pfau, D.; Spencer, J. S.; de G. Matthews, A. G.; Foulkes, W. (86) Zhou, Y.; Wu, J.; Chen, S.; Chen, G. To- M. C. Ab-Initio Solution of the Many- ward the Exact ExchangeâĂŞCorrelation Electron Schrödinger Equation with Potential: A Three-Dimensional Convo- Deep Neural Networks. 2019. lutional Neural Network Construct. J. Phys. Chem. Lett. 2019, 10, 7264–7269. (78) Hermann, J.; Schätzle, Z.; Noé, F. Deep Neural Network Solution of the Elec- (87) Kolb, B.; Lentz, L. C.; Kolpak, A. M. tronic Schrödinger Equation. 2019. Discovering Charge Density Function- als and Structure-Property Relationships (79) Gastegger, M.; McSloy, A.; Luya, M.; with PROPhet: A General Frame- Schütt, K. T.; Maurer, R. J. A work for Coupling Machine Learning Deep Neural Network for Molec- and First-Principles Methods. Sci. Rep. ular Wave Functions in Quasi- 2017, 7 . Atomic Minimal Basis Representation. https://arxiv.org/abs/2005.06979 2020, (88) Willatt, M. J.; Musil, F.; Ceriotti, M. Atom-Density Representations for Ma- (80) Hegde, G.; Bowen, R. C. Machine- chine Learning. J. Chem. Phys. 2019, Learned Approximations to Density 150, 154110. Functional Theory Hamiltonians. Sci. Rep. 2017, 7, 42669. (89) Choo, K.; Carleo, G.; Regnault, N.; Ne- upert, T. Symmetries and Many-Body (81) Brockherde, F.; Vogt, L.; Li, L.; Tucker- Excitations with Neural-Network Quan- man, M. E.; Burke, K.; Müller, K.-R. By- tum States. Phys. Rev. Lett. 2018, 121, passing the Kohn-Sham Equations with 167204. Machine Learning. Nat. Commun. 2017, 8, 872. (90) Guan, Y.; Yarkony, D. R. Accurate Neu- ral Network Representation of the Ab (82) Gastegger, M.; González, L.; Mar- Initio Determined SpinâĂŞOrbit Inter- quetand, P. Exploring Density Func- action in the Diabatic Representation In- tional Subspaces with Genetic Algo- cluding the Effects of Conical Intersec- rithms. Monatsh. Chem. 2019, 150, 173– tions. J. Phys. Chem. Lett. 2020, 11, 182. 1848–1858.

57 (91) Carbogno, C.; Behler, J.; Reuter, K.; Regression. J. Chem. Phys. 2019, 150, Groß, A. Signatures of Nonadiabatic O2 041101. Dissociation at Al(111): First-Principles Fewest-Switches Study. Phys. Rev. B (98) Hobday, S.; Smith, R.; Belbruno, J. Ap- 2010, 81, 035410. plications of Neural Networks to Fitting Interatomic Potential Functions. Modell. (92) Westermayr, J.; Gastegger, M.; Simul. Mater. Sci. Eng. 1999, 7, 397. Menger, M. F. S. J.; Mai, S.; González, L.; Marquetand, P. Ma- (99) Bartók, A. P.; Payne, M. C.; Kondor, R.; chine Learning Enables Long Time Scale Csányi, G. Gaussian Approximation Po- Molecular Photodynamics Simulations. tentials: The Accuracy of Quantum Me- Chem. Sci. 2019, 10, 8100–8107. chanics, without the Electrons. Phys. Rev. Lett. 2010, 104, 136403. (93) Westermayr, J.; Faber, F. A.; Chris- tensen, A. S.; von Lilienfeld, O. A.; Mar- (100) Rupp, M.; Tkatchenko, A.; Müller, K.- quetand, P. Neural Networks and Ker- R.; von Lilienfeld, O. A. Fast and Accu- nel Ridge Regression for Excited States rate Modeling of Molecular Atomization + Energies with Machine Learning. Phys. Dynamics of CH2NH2 : From Single- State to Multi-State Representations and Rev. Lett. 2012, 108, 058301. Multi-Property Machine Learning Mod- (101) Li, Z.; Kermode, J. R.; De Vita, A. els. Mach. Learn.: Sci. Technol. 2020, 1, Molecular Dynamics with On-the- 025009. Fly Machine Learning of Quantum- (94) Shen, Y.; Yarkony, D. R. Construction Mechanical Forces. Phys. Rev. Lett. of Quasi-diabatic Hamiltonians That Ac- 2015, 114, 096405. curately Represent Ab Initio Determined (102) von Lilienfeld, O. A.; Ramakrishnan, R.; Adiabatic Electronic States Coupled by Rupp, M.; Knoll, A. Fourier Series of Conical Intersections for Systems on the Atomic Radial Distribution Functions: Order of 15 Atoms. Application to Cy- A Molecular Fingerprint for Machine clopentoxide Photoelectron Detachment Learning Models of Quantum Chemi- in the Full 39 Degrees of Freedom. J. cal Properties. Int. J. Quantum Chem. Phys. Chem. A 2020, 124, 4539–4548. 2015, 115, 1084–1093.

(95) Zhang, Y.; Ye, S.; Zhang, J.; Jiang, J.; (103) Gastegger, M.; Marquetand, P. High- Jiang, B. Efficient and Accurate Spec- Dimensional Neural Network Potentials troscopic Simulations with Symmetry- for Organic Reactions and an Improved Preserving Neural Network Models Training Algorithm. J. Chem. Theory for Tensorial Properties. arXiv 2020, Comput. 2015, 11, 2187–2198. 2004.13605 . (104) Rupp, M.; Ramakrishnan, R.; von Lilien- (96) Carbogno, C.; Behler, J.; Groß, A.; feld, O. A. Machine Learning for Quan- Reuter, K. Fingerprints for Spin- tum Mechanical Properties of Atoms in Selection Rules in the Interaction Molecules. J. Phys. Chem. Lett. 2015, 6, Dynamics of O2 at Al(111). Phys. Rev. 3309–3313. Lett. 2008, 101, 096104. (105) Behler, J. Perspective: Machine Learning (97) Polyak, I.; Richings, G. W.; Haber- Potentials for Atomistic Simulations. J. shon, S.; Knowles, P. J. Direct Quan- Chem. Phys. 2016, 145, 170901. tum Dynamics using Variational Gaus- sian Wavepackets and Gaussian Process

58 (106) Artrith, N.; Urban, A. An implemen- Gaussian Approximation Potential Mod- tation of artificial neural-network po- eling of Lithium Intercalation in Carbon tentials for atomistic materials simula- Nanostructures. J. Chem. Phys. 2018, tions: Performance for TiO2. computa- 148, 241714. tional Materials Science 2016, 114, 135 – 150. (115) Behler, J. First Principles Neural Net- work Potentials for Reactive Simulations (107) Gastegger, M.; Kauffmann, C.; of Large Molecular and Condensed Sys- Behler, J.; Marquetand, P. Compar- tems. Angew. Chem. Int. Edit. 2017, 56, ing the Accuracy of High-Dimensional 12828–12840. Neural Network Potentials and the Systematic Molecular Fragmentation (116) Zong, H.; Pilania, G.; Ding, X.; Ack- Method: A Benchmark Study for All- land, G. J.; Lookman, T. Developing Trans Alkanes. J. Chem. Phys. 2016, an Interatomic Potential for Martensitic 144 . Phase Transformations in Zirconium by Machine Learning. npj comput Mater (108) Artrith, N.; Urban, A.; Ceder, G. Effi- 2018, 4 . cient and Accurate Machine-Learning In- terpolation of Atomic Energies in Com- (117) Wood, M. A.; Thompson, A. P. Extend- positions with Many Species. Phys. Rev. ing the Accuracy of the SNAP Inter- B 2017, 96, 014112. atomic Potential Form. J. Chem. Phys. 2018, 148, 241721. (109) Gastegger, M.; Behler, J.; Marque- tand, P. Machine Learning Molecular Dy- (118) Chen, X.; Jørgensen, M. S.; Li, J.; Ham- namics for the Simulation of Infrared mer, B. Atomic Energies from a Convolu- Spectra. Chem. Sci. 2017, 8, 6924–6935. tional Neural Network. J. Chem. Theory Comput. 2018, 14, 3933–3942. (110) Deringer, V. L.; Csányi, G. Machine Learning Based Interatomic Potential for (119) Bartók, A. P.; Kermode, J.; Bern- Amorphous Carbon. Phys. Rev. B: Con- stein, N.; Csányi, G. Machine Learn- dens. Matter Mater. Phys. 2017, 95, ing a General-Purpose Interatomic Po- 094203. tential for Silicon. Phys. Rev. X 2018, 8, 041048. (111) Botu, V.; Batra, R.; Chapman, J.; Ramprasad, R. Machine Learning Force (120) Chmiela, S.; Sauceda, H. E.; Müller, K.- Fields: Construction, Validation, and R.; Tkatchenko, A. Towards Exact Outlook. J. Phys. Chem. C 2017, 121, Molecular Dynamics Simulations with 511–522. Machine-Learned Force Fields. Nat. Commun. 2018, 9, 3887. (112) Glielmo, A.; Sollich, P.; De Vita, A. Ac- curate Interatomic Force Fields via Ma- (121) Imbalzano, G.; Anelli, A.; Giofré, D.; chine Learning with Covariant Kernels. Klees, S.; Behler, J.; Ceriotti, M. Au- Phys. Rev. B 2017, 95, 214302. tomatic Selection of Atomic Finger- prints and Reference Configurations for (113) Smith, J. S.; Isayev, O.; Roitberg, A. E. Machine-Learning Potentials. J. Chem. ANI-1: An Extensible Neural Network Phys. 2018, 148, 241730. Potential with DFT Accuracy at Force Field Computational Cost. Chem. Sci. (122) Zhang, L.; Han, J.; Wang, H.; 2017, 8, 3192–3203. Saidi, W. A.; Car, R.; Weinan, E. End-to-end Symmetry Preserving Inter- (114) Fujikake, S.; Deringer, V. L.; Lee, T. H.; atomic Potential Energy Model for Finite Krynski, M.; Elliott, S. R.; Csányi, G. and Extended Systems. Proceedings of

59 the 32Nd International conference on Vogt-Maranto, L.; Zdeborová, L. Ma- Neural Information Processing Systems. chine Learning and the Physical Sciences. USA, 2018; pp 4441–4451. Rev. Mod. Phys. 2019, 91, 045002.

(123) Zhang, L.; Han, J.; Wang, H.; Car, R.; (130) Krems, R. V. Bayesian Machine Learning E, W. Deep Potential Molecular Dynam- for Quantum Molecular Dynamics. Phys. ics: A Scalable Model with the Accuracy Chem. Chem. Phys. 2019, 21, 13392– of Quantum Mechanics. Phys. Rev. Lett. 13410. 2018, 120, 143001. (131) Deringer, V. L.; Caro, M. A.; Csányi, G. (124) Chan, H.; Narayanan, B.; Machine Learning Interatomic Potentials Cherukara, M. J.; Sen, F. G.; Sasiku- as Emerging Tools for Materials Science. mar, K.; Gray, S. K.; Chan, M. K. Y.; Adv. Mat. 2019, 31, 1902765. Sankaranarayanan, S. K. R. S. Ma- (132) Ward, L.; Blaiszik, B.; Foster, I.; As- chine Learning Classical Interatomic sary, R. S.; Narayanan, B.; Curtiss, L. Potentials for Molecular Dynamics from Machine Learning Prediction of Accu- First-Principles Training Data. J. Phys. rate Atomization Energies of Organic Chem. C 2019, 123, 6941–6957. Molecules from Low-Fidelity Quantum (125) Faber, F. A.; Christensen, A. S.; Chemical Calculations. MRS Commun. Huang, B.; von Lilienfeld, O. A. Alchem- 2019, 9, 891âĂŞ899. ical and Structural Distribution Based (133) Noé, F.; Tkatchenko, A.; Müller, K.- Representation for Universal Quantum R.; Clementi, C. Machine Learning for Machine Learning. J. Chem. Phys. 2018, Molecular Simulation. Annu. Rev. Phys. 148, 241717. Chem. 2020, 71, 361–390. (126) Wang, H.; Yang, W. Toward Build- (134) Alborzpour, J. P.; Tew, D. P.; Haber- ing Protein Force Fields by Residue- shon, S. Efficient and Accurate Eval- Based Systematic Molecular Fragmenta- uation of Potential Energy Matrix El- tion and Neural Network. J. Chem. The- ements for Quantum Dynamics using ory Comput. 2019, 15, 1409–1417. Gaussian Process Regression. J. Chem. (127) Gerrits, N.; Shakouri, K.; Behler, J.; Phys. 2016, 145, 174112. Kroes, G.-J. Accurate Probabilities for (135) Cheng, Z.; Zhao, D.; Ma, J.; Li, W.; Li, S. Highly Activated Reaction of Polyatomic An On-the-Fly Approach to Construct Molecules on Surfaces Using a High- Generalized Energy-Based Fragmenta- Dimensional Neural Network Potential: tion Machine Learning Force Fields of CHD + Cu(111). J. Phys. Chem. Lett. 3 Complex Systems. J. Phys. Chem. A 2019, 10, 1763–1768. 2020, 124, 5007–5014. (128) Chmiela, S.; Sauceda, H. E.; (136) Behler, J.; Reuter, K.; Scheffler, M. Poltavsky, I.; MÃijller, K.-R.; Nonadiabatic Effects in the Dissociation Tkatchenko, A. sGDML: Constructing of Oxygen Molecules at the Al(111) Sur- Accurate and Data Efficient Molecular face. Phys. Rev. B 2008, 77, 115421. Force Fields using Machine Learning. Comput. Phys. Commun. 2019, 240, 38 (137) Hu, D.; Xie, Y.; Li, X.; Li, L.; – 45. Lan, Z. Inclusion of Machine Learning Kernel Ridge Regression Potential En- (129) Carleo, G.; Cirac, I.; Cranmer, K.; ergy Surfaces in On-the-Fly Nonadia- Daudet, L.; Schuld, M.; Tishby, N.; batic Molecular Dynamics Simulation. J. Phys. Chem. Lett. 2018, 9, 2725–2732.

60 (138) Dral, P. O.; Barbatti, M.; Thiel, W. (146) Wang, Y.; Xie, C.; Guo, H.; Nonadiabatic Excited-State Dynamics Yarkony, D. R. A Quasi-Diabatic with Machine Learning. J. Phys. Chem. Representation of the 1,21A States of Lett. 2018, 9, 5660–5663. Methylamine. J. Phys. Chem. A 2019, 123, 5231–5241. (139) Chen, W.-K.; Liu, X.-Y.; Fang, W.- H.; Dral, P. O.; Cui, G. Deep Learning (147) Richings, G. W.; Habershon, S. Di- for Nonadiabatic Excited-State Dynam- rect Grid-Based Quantum Dynamics on ics. J. Phys. Chem. Lett. 2018, 9, 6702– Propagated Diabatic Potential Energy 6708. Surfaces. Chem. Phys. Lett. 2017, 683, 228 – 233. (140) Williams, D. M. G.; Eisfeld, W. Neural Network Diabatization: A New Ansatz (148) Netzloff, H. M.; collins, M. A.; Gor- for Accurate High-Dimensional Coupled don, M. S. Growing Multiconfigurational Potential Energy Surfaces. J. Chem. Potential Energy Surfaces with Applica- Phys. 2018, 149, 204106. tions to X+H2 (X=C,N,O) Reactions. J. Chem. Phys. 2006, 124, 154104. (141) Xie, C.; Zhu, X.; Yarkony, D. R.; Guo, H. Permutation Invariant Polynomial Neu- (149) Bettens, R. P. A.; Collins, M. A. Learn- ral Network Approach to Fitting Po- ing to Interpolate Molecular Potential tential Energy Surfaces. IV. Coupled Energy Surfaces with Confidence: A Diabatic Potential Energy Matrices. J. Bayesian Approach. J. Chem. Phys. Chem. Phys. 2018, 149, 144107. 1999, 111, 816–826.

(142) Guan, Y.; Zhang, D. H.; Guo, H.; (150) Ghosh, K.; Stuke, A.; Todorović, M.; Yarkony, D. R. Representation of Cou- Jørgensen, P. B.; Schmidt, M. N.; Ve- pled Adiabatic Potential Energy Sur- htari, A.; Rinke, P. Deep Learning Spec- faces using Neural Network Based Quasi- troscopy: Neural Networks for Molecu- Diabatic Hamiltonians: 1,2 2A’ States of lar Excitation Spectra. Adv. Sci. 2019, LiFH. Phys. Chem. Chem. Phys. 2019, 6, 1801367. 10.1039/C8CP06598E. (151) Kananenka, A. A.; Yao, K.; Cor- (143) Richings, G. W.; Habershon, S. MCTDH celli, S. A.; Skinner, J. L. Machine Learn- on-the-Fly: Efficient Grid-Based Quan- ing for Vibrational Spectroscopic Maps. tum Dynamics without Pre-Computed J. Chem. Theory Comput. 2019, 15, Potential Energy Surfaces. J. Chem. 6850–6858. Phys. 2018, 148, 134116. (152) Roch, L. M.; Saikin, S. K.; HÃďse, F.; (144) Richings, G. W.; Robertson, C.; Haber- Friederich, P.; Goldsmith, R. H.; shon, S. Improved on-the-Fly MCTDH León, S.; Aspuru-Guzik, A. From Ab- Simulations with Many-Body-Potential sorption Spectra to Charge Transfer Tensor Decomposition and Projection in Nanoaggregates of Oligomers with Diabatization. J. Chem. Theory Comput. Machine Learning. ACS Nano 2020, in 2019, 15, 857–870. press, doi:10.1021/acsnano.0c00384.

(145) Guan, Y.; Guo, H.; Yarkony, D. R. (153) Rankine, C. D.; Madkhali, M. M. M.; Neural Network Based Quasi-Diabatic Penfold, T. J. A Deep Neural Network for Hamiltonians with Symmetry Adapta- the Rapid Prediction of X-Ray Absorp- tion and a Correct Description of Conical tion Spectra. J. Phys. Chem. A 2020, Intersections. J. Chem. Phys. 2019, 150, 124, 4263–4270. 214101.

61 (154) Pereira, F.; Xiao, K.; Latino, D. A. R. S.; (162) Teunissen, J. L.; De Proft, F.; Wu, C.; Zhang, Q.; Aires-de Sousa, J. De Vleeschouwer, F. Tuning the Machine Learning Methods to Predict HOMOâĂŞLUMO Energy Gap of Small Density Functional Theory B3LYP En- Diamondoids Using Inverse Molecular ergies of HOMO and LUMO Orbitals. J. Design. J. Chem. Theory Comput. 2017, Chem. Inf. Model. 2017, 57, 11–21. 13, 1351–1365. (155) Isayev, O.; Oses, c.; Toher, c.; Gos- (163) Liu, D.; Tan, Y.; Khoram, E.; Yu, Z. sett, E.; Curtarolo, S.; Tropsha, A. Uni- Training Deep Neural Networks for the versal Fragment Descriptors for Predict- Inverse Design of Nanophotonic Struc- ing Properties of Inorganic Crystals. Nat. tures. ACS Photonics 2018, 5, 1365– Commun. 2017, 8, 15679. 1369. (156) Pronobis, W.; Schütt, K. R.; (164) Elton, D. C.; Boukouvalas, Z.; Tkatchenko, A.; Müller, K.-R. Capturing Fuge, M. D.; Chung, P. W. Deep Intensive and Extensive DFT/TDDFT Learning for Molecular Design – A Molecular Properties with Machine Review of the State of the Art. Mol. Learning. Eur. Phys. J. B 2018, 91, Syst. Des. Eng. 2019, 4, 828–849. 178. (165) Sanchez-Lengeling, B.; Aspuru-Guzik, A. (157) Stuke, A.; Todorović, M.; Rupp, M.; Inverse Molecular Design using Machine Kunkel, C.; Ghosh, K.; Himanen, L.; Learning: Generative Models for Mat- Rinke, P. Chemical Diversity in Molecu- ter Engineering. Science 2018, 361, 360– lar Orbital Energy Predictions with Ker- 365. nel Ridge Regression. J. Chem. Phys. 2019, 150, 204121. (166) Goldsmith, B. R.; Esterhuizen, J.; Liu, J.-X.; Bartel, C. J.; Sutton, C. Ma- (158) Häse, F.; Fdez. Galván, I.; Aspuru- chine Learning for Heterogeneous Cat- Guzik, A.; Lindh, R.; Vacher, M. How alyst Design and Discovery. AIChE J. Machine Learning can Assist the Inter- 2018, 64, 2311–2323. pretation of Ab Initio Molecular Dynam- ics Simulations and Conceptual Under- (167) Davies, D. W.; Butler, K. T.; Isayev, O.; standing of Chemistry. Chem. Sci. 2019, Walsh, A. Materials Discovery by Chem- 10, 2298–2307. ical Analogy: Role of Oxidation States in Structure Prediction. Faraday Discuss. (159) Häse, Florian and Kreisbeck, Christoph 2018, 211, 553–568. and Aspuru-Guzik, Alán, Machine Learning for Quantum Dynamics: Deep (168) Anatole von Lilienfeld, O.; Müller, K.-R.; Learning of Excitation Energy Trans- Tkatchenko, A. Exploring Chemical com- fer Properties. Chem. Sci. 2017, 8, pound Space with Quantum-Based Ma- 8419–8426. chine Learning. Nat. Rev. Chem. 2020, (160) Häse, F.; Valleau, S.; Pyzer-Knapp, E.; (169) Freeze, J. G.; Kelly, H. R.; Batista, V. S. Aspuru-Guzik, A. Machine Learning Ex- Search for Catalysts by Inverse De- citon Dynamics. Chem. Sci. 2016, 7, sign: Artificial Intelligence, Mountain 5139–5147. Climbers, and Alchemists. Chemical Re- views 2019, 119, 6595–6612. (161) OâĂŹBoyle, N. M.; Campbell, C. M.; Hutchison, G. R. Computational Design (170) Lee, M.-H. Robust Random Forest Based and Selection of Optimal Organic Pho- Non-Fullerene Organic Solar Cells Effi- tovoltaic Materials. J. Phys. Chem. C ciency Prediction. Org. Electron. 2020, 2011, 115, 16200–16210. 76, 105465.

62 (171) Cartwright, H. M., Ed. Machine Learn- (180) Tavernelli, I. Electronic Density Re- ing in Chemistry; Theoretical and Com- sponse of Liquid Water using Time- putational Chemistry Series; The Royal Dependent Density Functional Theory. Society of Chemistry, 2020. Phys. Rev. B 2006, 73, 094204.

(172) Gastegger, M.; Marquetand, P. In Ma- (181) Schütt, K. T.; Glawe, H.; Brockherde, F.; chine Learning Meets Quantum Physics; Sanna, A.; Müller, K. R.; Gross, E. K. U. Schütt, K. T., Chmiela, S., von Lilien- How to Represent Crystal Structures for feld, O. A., Tkatchenko, A., Tsuda, K., Machine Learning: Towards Fast Predic- Müller, K.-R., Eds.; Springer Interna- tion of Electronic Properties. Phys. Rev. tional Publishing: Cham, 2020; pp 233– B 2014, 89, 205118. 252. (182) Lee, J.; Seko, A.; Shitara, K.; (173) Schütt, K. T., Chmiela, S., von Lilien- Nakayama, K.; Tanaka, I. Predic- feld, O. A., Tkatchenko, A., Tsuda, K., tion Model of Band Gap for Inorganic Müller, K.-R., Eds. Machine Learning Compounds by Combination of Density Meets Quantum Physics; Springer Inter- Functional Theory Calculations and national Publishing, 2020. Machine Learning Techniques. Phys. Rev. B 2016, 93, 115104. (174) Park, J. W.; Al-Saadon, R.; MacLeod, M. K.; Shiozaki, T.; Vlaisavl- (183) Zhuo, Y.; Mansouri Tehrani, A.; Br- jevich, B. Multireference Electron goch, J. Predicting the Band Gaps of In- Correlation Methods: Journeys along organic Solids by Machine Learning. J. Potential Energy Surfaces. Chem. Rev. Phys. Chem. Lett. 2018, 9, 1668–1673. 2020, in press, null. (184) Pilania, G.; Gubernatis, J.; Lookman, T. (175) Akimov, A. V.; Prezhdo, O. V. Large- Multi-Fidelity Machine Learning Mod- Scale Computations in Chemistry: A els for Accurate Bandgap Predictions of Bird’s Eye View of a Vibrant Field. Solids. Comput. Mat. Sci. 2017, 129, 156 Chem. Rev. 2015, 115, 5797–5890. – 163.

(176) Frutos, L. M.; Andruniów, T.; San- (185) Spiering, P.; Shakouri, K.; Behler, J.; toro, F.; Ferré, N.; Olivucci, M. Track- Kroes, G.-J.; Meyer, J. Orbital- ing the Excited-State Time Evolution of Dependent Electronic Friction Sig- the Visual Pigment with Multiconfigu- nificantly Affects the Description rational Quantum Chemistry. Proceed- of Reactive Scattering of N2 from ings of the National Academy of Sciences Ru(0001). J. Phys. Chem. Lett. 2019, 2007, 104, 7764–7769. 10, 2957–2962.

(177) Menger, M. F. S. J.; Plasser, F.; Men- (186) Zhang, Y.; Maurer, R. J.; Jiang, B. nucci, B.; González, L. Surface Hop- Symmetry-Adapted High Dimensional ping within an Exciton Picture. An Elec- Neural Network Representation of Elec- trostatic Embedding Scheme. J. Chem. tronic Friction Tensor of Adsorbates on Theory Comput. 2018, 14, 6139–6148. Metals. J. Phys. Chem. C 2020, 124, 186–195. (178) Dou, W.; Subotnik, J. E. Nonadiabatic Molecular Dynamics at Metal Surfaces. (187) Zhang, Y.; Maurer, R. J.; Guo, H.; J. Phys. Chem. A 2020, 124, 757–771. Jiang, B. Hot-Electron Effects during Re- (179) Dou, W.; Nitzan, A.; Subotnik, J. E. active Scattering of H2 from Ag(111): Frictional Effects Near a Metal Surface. The Interplay between Mode-Specific J. Chem. Phys. 2015, 143, 054103.

63 Electronic Friction and the Potential En- (196) Shenvi, N.; Roy, S.; Tully, J. C. Nona- ergy Landscape. Chem. Sci. 2019, 10, diabatic Dynamics at Metal Surfaces: 1089–1097. Independent-Electron Surface Hopping. J. Chem. Phys. 2009, 130, 174107. (188) Head-Gordon, M.; Tully, J. C. Molecu- lar Dynamics with Electronic Frictions. (197) Shenvi, N.; Roy, S.; Tully, J. C. Dynami- J. Chem. Phys. 1995, 103, 10137–10145. cal Steering and Electronic Excitation in NO Scattering from a Gold Surface. Sci- (189) Douglas-Gallardo, O. A.; Berdakin, M.; ence 2009, 326, 829–832. Frauenheim, T.; Sánchez, C. G. Plasmon-Induced Hot-Carrier Gen- (198) Dou, W.; Schinabeck, C.; Thoss, M.; eration Differences in Gold and Silver Subotnik, J. E. A broadened classi- Nanoclusters. Nanoscale 2019, 11, cal master equation approach for treat- 8604–8615. ing electron-nuclear coupling in non- equilibrium transport. J. Chem. Phys. (190) Yin, R.; Zhang, Y.; Jiang, B. Strong 2018, 148, 102317. Vibrational Relaxation of NO Scattered from Au(111): Importance of the Adia- (199) Jiang, B.; Li, J.; Guo, H. High-Fidelity batic Potential Energy Surface. J. Phys. Potential Energy Surfaces for Gas Phase Chem. Lett. 2019, 10, 5969–5974. and Gas-Surface Scattering Processes from Machine Learning. J. Phys. Chem. (191) Rittmeyer, S. P.; Bukas, V. J.; Reuter, K. Lett. 2020, 11, 5120âĂŞ5131. Energy dissipation at metal surfaces. Adv. Phys-X 2018, 3, 1381574. (200) Buhrke, D.; Hildebrandt, P. Probing Structure and Reaction Dynamics of Pro- (192) Therrien, A. J.; Kale, M. J.; Yuan, L.; teins Using Time-Resolved Resonance Zhang, C.; Halas, N. J.; Christopher, P. Raman Spectroscopy. Chem. Rev. 2020, Impact of Chemical Interface Damping 120, 3577–3630. on Surface Plasmon Dephasing. Faraday Discuss. 2019, 214, 59–72. (201) Raimbault, N.; Grisafi, A.; Ceriotti, M.; Rossi, M. Using Gaussian Process Re- (193) Wodtke, A. M.; Tully, J. C.; Auer- gression to Simulate the Vibrational Ra- bach, D. J. Electronically Non-Adiabatic man Spectra of Molecular Crystals. New Interactions of Molecules at Metal Sur- J. Phys. 2019, 21, 105001. faces: Can we Trust the BornâĂŞOppen- heimer Approximation for Surface Chem- (202) Hu, W.; Ye, S.; Zhang, Y.; Li, T.; istry? Int. Rev. Phys. Chem. 2004, 23, Zhang, G.; Luo, Y.; Mukamel, S.; 513–539. Jiang, J. Machine Learning Protocol for Surface-Enhanced Raman Spectroscopy. (194) Park, G. B.; KrÃijger, B. C.; J. Phys. Chem. Lett. 2019, 10, 6026– Borodin, D.; Kitsopoulos, T. N.; 6031. Wodtke, A. M. Fundamental Mecha- nisms for Molecular Energy Conversion (203) Lussier, F.; Thibault, V.; Charron, B.; and Chemical Reactions at Surfaces. Wallace, G. Q.; Masson, J.-F. Deep Rep. Prog. Phys. 2019, 82, 096401. Learning and Artificial Intelligence Methods for Raman and Surface- (195) Jiang, B.; Guo, H. Dynamics in Reac- Enhanced Raman Scattering. TrAC, tions on Metal Surfaces: A Theoretical Trends Anal. Chem. 2020, 124, 115796. Perspective. J. Chem. Phys. 2019, 150, 180901. (204) Fu, W.; Hopkins, W. S. Applying Ma- chine Learning to Vibrational Spec-

64 troscopy. J. Phys. Chem. A 2018, 122, (213) Öhlknecht, C.; Lier, B.; Petrov, D.; 167–171. Fuchs, J.; Oostenbrink, C. Correct- ing Electrostatic Artifacts due to Net- (205) Aires-de Sousa, J.; Hemmer, M. C.; Charge Changes in the Calculation of 1 Gasteiger, J. Prediction of H NMR Ligand Binding Free Energies. J. Com- Chemical Shifts Using Neural Networks. put. Chem. 2020, 41, 986–999. Anal. Chem. 2002, 74, 80–90. (214) Michlits, H.; Lier, B.; Pfanzagl, V.; (206) Taguchi, A. T.; Evans, E. D.; Djinović-Carugo, K.; Furtmüller, P. G.; Dikanov, S. A.; Griffin, R. G. Con- Oostenbrink, C.; Obinger, C.; Hof- volutional Neural Network Analysis of bauer, S. Actinobacterial Coproheme De- Two-Dimensional Hyperfine Sublevel carboxylases Use Histidine as a Distal Correlation Electron Paramagnetic Base to Promote Compound I Forma- Resonance Spectra. J. Phys. Chem. Lett. tion. ACS Catal. 2020, 10, 5405–5418. 2019, 10, 1115–1119. (215) Brunk, E.; Rothlisberger, U. Mixed (207) Cobas, C. NMR Signal Processing, Pre- Quantum Mechanical/Molecular Me- diction, and Structure Verification with chanical Molecular Dynamics Simula- Machine Learning Techniques. Magn. tions of Biological Systems in Ground Reson. Chem. 2020, 58, 512–519. and Electronically Excited States. Chem. (208) Salomon-Ferrer, R.; Case, D. A.; Rev. 2015, 115, 6217–6263. Walker, R. C. An overview of the Amber (216) Bedrov, D.; Piquemal, J.-P.; Borodin, O.; biomolecular simulation package. WIREs MacKerell, A. D.; Roux, B.; Schröder, C. Computational Molecular Science 2013, Molecular Dynamics Simulations of Ionic 3, 198–210. Liquids and Electrolytes Using Polariz- (209) B. r. Brooks, et al., CHARMM: The able Force Fields. Chem. Rev. 2019, 119, Biomolecular Simulation Program. J. 7940–7995. Comput. Chem. 2009, 30, 1545–1614. (217) Sosso, G. C.; Chen, J.; Cox, S. J.; (210) Eichenberger, A. P.; Allison, J. R.; Fitzner, M.; Pedevilla, P.; Zen, A.; Dolenc, J.; Geerke, D. P.; Horta, B. Michaelides, A. Crystal Nucleation in A. C.; Meier, K.; Oostenbrink, C.; Liquids: Open Questions and Future Schmid, N.; Steiner, D.; Wang, D.; van Challenges in Molecular Dynamics Sim- Gunsteren, W. F. GROMOS++ Soft- ulations. Chem. Rev. 2016, 116, 7078– ware for the Analysis of Biomolecular 7116. Simulation Trajectories. J. Chem. The- (218) Venable, R. M.; Krämer, A.; Pas- ory Comput. 2011, 7, 3379–3390. tor, R. W. Molecular Dynamics Simula- (211) Reif, M. M.; HÃijnenberger, P. H.; Oost- tions of Membrane Permeability. Chem. enbrink, C. New Interaction Parameters Rev. 2019, 119, 5954–5997. for Charged Amino Acid Side Chains (219) Marrink, S. J.; Corradi, V.; Souza, P. C.; in the GROMOS Force Field. J. Chem. IngÃşlfsson, H. I.; Tieleman, D. P.; San- Theory Comput. 2012, 8, 3705–3723. som, M. S. Computational Modeling of (212) Perthold, J. W.; Petrov, D.; Oost- Realistic Cell Membranes. Chemical Re- enbrink, C. Towards Automated Free views 2019, 119, 6184–6226. Energy Calculation with Accelerated Enveloping Distribution Sampling (A- (220) G., G. In Biomolecular Simulations. EDS). J. Chem. Inf. Model. 2020, in Methods in Molecular Biology (Meth- press, doi:10.1021/acs.jcim.0c00456. ods and Protocols); Monticelli, L., Salo-

65 nen, E., Eds.; Humana Press, Totowa, Physics; Schütt, K. T., Chmiela, S., NJ, 2013; Vol. 924. von Lilienfeld, O. A., Tkatchenko, A., Tsuda, K., Müller, K.-R., Eds.; Springer (221) Thomas P. Senftle, et al., The ReaxFF International Publishing: Cham, 2020; Reactive Force-Field: Development, Ap- pp 171–194. plications and Future Directions. npj Comput. Mater. 2016, 2 . (228) Köppel, H.; Domcke, W.; Ceder- baum, L. S. in: Conical Intersections (222) Sauceda, H. E.; Chmiela, S.; (W. Domcke, D. R. Yarkony, H. Köp- Poltavsky, I.; Müller, K.-R.; pel, Eds.); World Scientific: New York, Tkatchenko, A. In Machine Learning 2004. Meets Quantum Physics; Schütt, K. T., Chmiela, S., von Lilienfeld, O. A., (229) Plasser, F.; GÃşmez, S.; Menger, M. F. Tkatchenko, A., Tsuda, K., Müller, K.- S. J.; Mai, S.; González, L. Highly Effi- R., Eds.; Springer International Publish- cient Surface Hopping Dynamics using a ing: Cham, 2020; pp 277–307. Linear Vibronic Coupling Model. Phys. Chem. Chem. Phys. 2019, 21, 57–69. (223) Noé, F. In Machine Learning Meets Quantum Physics; Schütt, K. T., (230) He, G. S.; Tan, L.-S.; Zheng, Q.; Chmiela, S., von Lilienfeld, O. A., Prasad, P. N. Multiphoton Absorbing Tkatchenko, A., Tsuda, K., Müller, K.- Materials: Molecular Designs, Char- R., Eds.; Springer International Publish- acterizations, and Applications. Chem. ing: Cham, 2020; pp 331–372. Rev. 2008, 108, 1245–1330.

(224) Glielmo, A.; Zeni, C.; Fekete, Á.; (231) Marquetand, P.; Weinacht, T.; Roz- De Vita, A. In Machine Learning gonyi, T.; González-Vazquez, J.; Meets Quantum Physics; Schütt, K. T., Geiçler, D.; González, L. In Ad- Chmiela, S., von Lilienfeld, O. A., vances in Multiphoton Processes and Tkatchenko, A., Tsuda, K., Müller, K.- Spectroscopy; Fujimura, Y., Ed.; World R., Eds.; Springer International Publish- Scientific, Singapore, 2014; Vol. 21; pp ing: Cham, 2020; pp 67–98. 1–54.

(225) Abbott, A. S.; Turney, J. M.; Zhang, B.; (232) Tagliamonti, V.; Sándor, P.; Zhao, A.; Smith, D. G. A.; Altarawy, D.; Schae- Rozgonyi, T.; Marquetand, P.; fer, H. F. PES-Learn: An Open-Source Weinacht, T. Nonadiabatic Dynam- Software Package for the Automated ics and Multiphoton Resonances in Generation of Machine Learning Mod- Strong-Field Molecular Ionization with els of Molecular Potential Energy Sur- Few-Cycle Laser Pulses. Phys. Rev. A faces. J. Chem. Theory Comput. 2019, 2016, 93, 051401. 15, 4386–4398. (233) M. Wollenhaupt, A. A.; Baumert, T. In (226) Hellström, M.; Behler, J. In Ma- Springer Handbook of Lasers and Optics; chine Learning Meets Quantum Physics; Träger, F., Ed.; Springer Science and Schütt, K. T., Chmiela, S., von Lilien- Business Media, LLC New York, 2007; feld, O. A., Tkatchenko, A., Tsuda, K., Chapter 12, pp 937–983. Müller, K.-R., Eds.; Springer Interna- tional Publishing: Cham, 2020; pp 253– (234) Hilborn, R. C. Einstein Coefficients, 275. Cross Sections, f Values, Dipole Mo- ments, and All That. Am. J. Phys. 1982, (227) Vargas-Hernández, R. A.; Krems, R. V. 50, 982–986. In Machine Learning Meets Quantum

66 (235) Andrews, D. L. Molecular Photophysics Theory. Annu. Rev. Phys. Chem. 2012, and Spectroscopy; 2053-2571; Morgan & 63, 287–323. Claypool Publishers, 2014; pp 9–1 to 9–4. (245) Maitra, N. T. Perspective: Fundamen- (236) Silva, G. L.; Ediz, V.; Yaron, D.; Ar- tal Aspects of Time-Dependent Den- mitage, B. A. Experimental and Com- sity Functional Theory. J. Chem. Phys. putational Investigation of Unsymmetri- 2016, 144, 220901. cal Cyanine Dyes: Understanding Tor- sionally Fluorogenic Dyes. J. Am. Chem. (246) Szalay, P. G.; Müller, T.; Gidofalvi, G.; Soc. 2007, 129, 5710–5718. Lischka, H.; Shepard, R. Multiconfigura- tion Self-Consistent Field and Multirefer- (237) Hartschuh, A.; Pedrosa, H. N.; ence Configuration Interaction Methods Novotny, L.; Krauss, T. D. Simultaneous and Applications. Chem. Rev. 2012, 112, Fluorescence and Raman Scattering 108–181. from Single Carbon Nanotubes. Science 2003, 301, 1354–1356. (247) Helgaker, T.; Jørgensen, P.; Olsen, J. Molecular ElectronicâĂŘStructure The- (238) Terenziani, F.; Katan, C.; Badaeva, E.; ory; John Wiley & Sons, Ltd, 2014. Tretiak, S.; Blanchard-Desce, M. En- hanced Two-Photon Absorption of Or- (248) Roos, B. O.; Lindh, R.; Malmqvist, P. Å.; ganic Chromophores: Theoretical and Veryazov, V.; Widmark, P. Multiconfigu- Experimental Assessments. Adv. Mater. rational Quantum Chemistry; John Wi- 2008, 20, 4641–4678. ley & Sons, Ltd, 2016.

(239) Richings, G. W.; Habershon, S. A New (249) Born, M.; Oppenheimer, R. Zur Quan- Diabatization Scheme for Direct Quan- tentheorie der Molekeln. Ann. Phys. tum Dynamics: Procrustes Diabatiza- 1927, 389, 457–484. tion. J. Chem. Phys. 2020, 152, 154108. (250) Kohn, W. Nobel Lecture: Electronic (240) Tannor, D. Introduction to Quantum Me- Structure of Matter – Wave Functions chanics: A Time-Dependent Perspec- and Density Functionals. Rev. Mod. tive; University Science Books: Sausal- Phys. 1999, 71, 1253–1266. ito, 2006. (251) Schrödinger, E. An Undulatory The- (241) Weinacht, T.; Pearson, B. Time-Resolved ory of the Mechanics of Atoms and Spectroscopy: An Experimental Perspec- Molecules. Phys. Rev. 1926, 28, 1049– tive; CRC Press: New York, 2019. 1070.

(242) Mai, S.; Marquetand, P.; González, L. (252) Erwin-Schrödinger – Nobel Lecture. Nonadiabatic Dynamics: The SHARC https://www.nobelprize.org/prizes/phy- Approach. WIREs Comput. Mol. Sci. sics/1933/schrodinger/lecture/. 2018, 8, e1370. (253) Yu, H. S.; Li, S. L.; Truhlar, D. G. Per- (243) Yonehara, T.; Hanasaki, K.; Takat- spective: Kohn-Sham Density Functional suka, K. Fundamental Approaches to Theory Descending a Staircase. J. Chem. Nonadiabaticity: Toward a Chemical Phys. 2016, 145, 130901. Theory beyond the Born–Oppenheimer (254) Maurer, R. J.; Freysoldt, C.; Paradigm. Chem. Rev. 2012, 112, 499– Reilly, A. M.; Brandenburg, J. G.; 542. Hofmann, O. T.; BjÃűrkman, T.; (244) Casida, M.; Huix-Rotllant, M. Progress LebÃĺgue, S.; Tkatchenko, A. Advances in Time-Dependent Density-Functional in Density-Functional Calculations for

67 Materials Modeling. Annual Review of (264) Möller, C.; Plesset, M. S. Note on Materials Research 2019, 49, 1–30. an Approximation Treatment for Many- Electron Systems. Phys. Rev. 1934, 46, (255) Benavides-Riveros, C. L.; Lathio- 618–622. takis, N. N.; Marques, M. A. L. Towards a Formal Definition of Static (265) Bartlett, R. J. Many-Body Perturbation and Dynamic Electronic Correlations. Theory and Coupled Cluster Theory for Phys. Chem. Chem. Phys. 2017, 19, Electron Correlation in Molecules. Annu. 12655–12664. Rev. Phys. Chem. 1981, 32, 359–401.

(256) Szabo, A.; Ostlund, N. Modern Quantum (266) Helgaker, T.; Jørgensen, P.; Olsen, J. Chemistry: Introduction to Advanced Molecular ElectronicâĂŘStructure The- Electronic Structure Theory; Dover ory; John Wiley & Sons, Ltd, 2014; Books on Chemistry; Dover Publications, Chapter 13, pp 648–723. 2012. (267) Izsák, R. Single-Reference Coupled Clus- (257) Helgaker, T.; Jørgensen, P.; Olsen, J. ter Methods for Computing Excitation Molecular ElectronicâĂŘStructure The- Energies in Large Molecules: The Effi- ory; John Wiley & Sons, Ltd, 2014; ciency and Accuracy of Approximations. Chapter 10, pp 433–522. WIREs Comput. Mol. Sci. 2020, 10, e1445. (258) Helgaker, T.; Jørgensen, P.; Olsen, J. Molecular ElectronicâĂŘStructure The- (268) Krylov, A. I. Equation-of-Motion ory; John Wiley & Sons, Ltd, 2014; Coupled-Cluster Methods for Open- Chapter 11, pp 523–597. Shell and Electronically Excited Species: The Hitchhiker’s Guide to Fock Space. (259) Dreuw, A.; Wormit, M. The algebraic di- Annu. Rev. Phys. Chem. 2008, 59, agrammatic construction scheme for the 433–462. polarization propagator for the calcula- tion of excited states. WIREs Comput. (269) Parrill, A.; Lipkowitz, K. Reviews in Mol. Sci. 2015, 5, 82–95. Computational Chemistry, Volume 31 ; Reviews in Computational Chemistry; (260) von Niessen, W.; Schirmer, J.; Ceder- Wiley, 2018. baum, L. Computational Methods for the One-Particle Green’s Function. Comp. (270) Pacifici L., L. A., Verdicchio M. In Com- Phys. Rep. 1984, 1, 57 – 125. putational Science and Its Applications âĂŞ ICCSA 2013 ; B. Murgante, et. al„ (261) Linderberg, J.; Öhrn, Y. Propagators Ed.; Springer, Berlin, Heidelberg, 2013; in Quantum Chemistry; John Wiley & Vol. 7971. Sons, Ltd, 2005; Chapter 2, pp 3–6. (271) Helgaker, T.; Jørgensen, P.; Olsen, J. (262) Melin, J.; Ayers, P.; Ortiz, J. The Molecular ElectronicâĂŘStructure The- Electron-Propagator Approach to Con- ory; John Wiley & Sons, Ltd, 2014; ceptual Density-Functional Theory. J. Chapter 12, pp 598–647. Chem. Sci. 2005, 117, 387–400. (272) Roos, B. O.; Taylor, P. R.; Sieg- (263) Corzo, H. H.; Ortiz, J. V. In Löwdin Vol- bahn, P. E. A Complete Active Space ume; Sabin, J. R., Brändas, E. J., Eds.; SCF Method (CASSCF) using a Density Advances in Quantum Chemistry; Aca- Matrix Formulated Super-CI Approach. demic Press, 2017; Vol. 74; pp 267 – 298. Chem. Phys. 1980, 48, 157–173.

68 (273) Roos, B. O.; Siegbahn, P. E. M. A Direct (282) Maitra, R.; Sinha, D.; Mukherjee, D. CI Method with a Multiconfigurational Unitary Group Adapted State-Specific Reference State. Int. J. Quantum Chem. Multi-Reference Coupled Cluster The- 1980, 17, 485–500. ory: Formulation and Pilot Numerical Applications. J. Chem. Phys. 2012, 137, (274) Lischka, H.; Dallos, M.; Szalay, P. G.; 024105. Yarkony, D. R.; Shepard, R. Ana- lytic Evaluation of Nonadiabatic cou- (283) Máşik, J.; Hubaç, I. In Multireference pling Terms at the MR-CI Level. I. Brillouin-Wigner Coupled-Cluster The- Formalism. J. Chem. Phys. 2004, 120, ory. Single-Root Approach.; Sabin, J. R., 7322–7329. Zerner, M. C., Brändas, E., Wilson, S., Maruani, J., Smeyers, Y., Grout, P., (275) Lischka, H. et al. The Generality of the McWeeny, R., Eds.; Advances in Quan- GUGA MRCI Approach in COLUMBUS tum Chemistry; Academic Press, 1998; for Treating Complex Quantum Chem- Vol. 31; pp 75 – 104. istry. J. Chem. Phys. 2020, 152, 134110. (284) Musiał, M.; Perera, A.; Bartlett, R. J. (276) Andersson, K.; Malmqvist, P. A.; Multireference Coupled-Cluster Theory: Roos, B. O.; Sadlej, A. J.; Wolinski, K. The Easy Way. J. Chem. Phys. 2011, Second-Order Perturbation Theory with 134, 114108. a CASSCF Reference Function. J. Phys. Chem. 1990, 94, 5483–5488. (285) Evangelista, F. A. Perspective: Multiref- erence Coupled Cluster Theories of Dy- (277) Andersson, K.; Malmqvist, P.; namical Electron Correlation. J. Chem. Roos, B. O. SecondâĂŘOrder Per- Phys. 2018, 149, 030901. turbation Theory with a Complete Active Space SelfâĂŘConsistent Field (286) Fdez. Galván, I. et al. OpenMolcas: From Reference Function. J. Phys. Chem. Source Code to Insight. J. Chem. Theory 1992, 96, 1218–1226. Comput. 2019, 15, 5925–5964.

(278) Finley, J.; Malmqvist, P.-A.; Roos, B. O.; (287) Roos, B.; Lindh, R.; Malmqvist, P.- Serrano-Andrés, L. The Multi-State Å.; Veryazov, V.; Widmark, P.-O. Main {CASPT2} Method. Chem. Phys. Lett. Group Atoms and Dimers Studied with a 1998, 288, 299 – 306. new Relativistic ANO Basis Set. J. Phys. Chem. A 2004, 108, 2851–2858. (279) Angeli, C.; Cimiraglia, R.; Evange- listi, S.; Leininger, T.; Malrieu, J.-P. In- (288) Vogiatzis, K. D.; Ma, D.; Olsen, J.; troduction of N-Electron Valence States Gagliardi, L.; de Jong, W. A. Pushing for Multireference Perturbation Theory. Configuration-Interaction to the limit: J. Chem. Phys. 2001, 114, 10252–10264. Towards Massively Parallel MCSCF Cal- culations. J. Chem. Phys. 2017, 147, (280) Roemelt, M.; Guo, S.; Chan, G. K.-L. 184111. A Projected Approximation to Strongly Contracted N-Electron Valence Pertur- (289) Kato, H.; Baba, M. Dynamics of Excited bation Theory for DMRG Wavefunc- Molecules: Predissociation. Chem. Rev. tions. J. Chem. Phys. 2016, 144, 204113. 1995, 95, 2311–2349.

(281) Guo, Y.; Sivalingam, K.; Valeev, E. F.; (290) Merer, A. J.; Mulliken, R. S. Ultraviolet Neese, F. Explicitly Correlated N- Spectra and Excited States of Ethylene Electron Valence State Perturbation and its Alkyl Derivatives. Chem. Rev. Theory (NEVPT2-F12). J. Chem. Phys. 1969, 69, 639–656. 2017, 147, 064110.

69 (291) Ashfold, M. N. R.; Langford, S. R. In The (300) Freitag, L.; Knecht, S.; Angeli, C.; Rei- Role of Rydberg States in Spectroscopy her, M. Multireference Perturbation The- and Photochemistry: Low and High Ry- ory with Cholesky Decomposition for the dberg States; Sándorfy, C., Ed.; Springer Density Matrix Renormalization Group. Netherlands: Dordrecht, 1999; pp 23–56. J. Chem. Theory Comput. 2017, 13, 451–459. (292) Merkt, F. Molecules in High Rydberg States. Annu. Rev. Phys. Chem. 1997, (301) Freitag, L.; Ma, Y.; Baiardi, A.; 48, 675–709. Knecht, S.; Reiher, M. Approximate Analytical Gradients and Nonadiabatic (293) Stein, C. J.; Reiher, M. Automated Selec- Couplings for the State-Average Den- tion of Active Orbital Spaces. J. Chem. sity Matrix Renormalization Group Self- Theory Comput. 2016, 12, 1760–1771. Consistent-Field Method. J. Chem. The- (294) Stein, C. J.; Reiher, M. Measuring Multi- ory Comput. 2019, 15, 6724–6737. Configurational Character by Orbital (302) Hohenberg, P.; Kohn, W. Inhomogeneous Entanglement. Mol. Phys. 2017, 115, Electron Gas. Phys. Rev. 1964, 136, 2110–2119. B864–B871.

(295) Stein, C. J.; Reiher, M. Automated Iden- (303) Kohn, W.; Sham, L. J. Self-Consistent tification of Relevant Frontier Orbitals Equations Including Exchange and Cor- for Chemical Compounds and Processes. relation Effects. Phys. Rev. 1965, 140, CHIMIA 2017, 71, 170–176. A1133–A1138.

(296) Chan, G. K.-L.; Van Voorhis, T. (304) Casida, M. E. Time-Dependent Density- Density-Matrix Renormalization-Group Functional Theory for Molecules and Algorithms with Nonorthogonal Orbitals Molecular Solids. J. Mol. Struc.- and Non-Hermitian Operators, and Ap- Theochem 2009, 914, 3 – 18. plications to Polyenes. J. Chem. Phys. 2005, 122, 204101. (305) Runge, E.; Gross, E. K. U. Density- Functional Theory for Time-Dependent (297) Zgid, D.; Nooijen, M. The Density Systems. Phys. Rev. Lett. 1984, 52, 997– Matrix Renormalization Group Self- 1000. Consistent Field Method: Orbital Opti- mization with the Density Matrix Renor- (306) Zangwill, A.; Soven, P. Density- malization Group Method in the Ac- Functional Approach to Local-Field tive Space. J. Chem. Phys. 2008, 128, Effects in Finite Systems: Photoabsorp- 144116. tion in the Rare Gases. Phys. Rev. A 1980, 21, 1561–1572. (298) Keller, S.; Dolfi, M.; Troyer, M.; Rei- her, M. An Efficient Matrix Product Op- (307) Chong, D. P. Recent Advances in Den- erator Representation of the Quantum sity Functional Methods; World Scien- Chemical Hamiltonian. J. Chem. Phys. tific, 1995. 2015, 143, 244118. (308) Tamm, I. Relativistic Interaction of El- (299) Knecht, S.; Keller, S.; Autschbach, J.; ementary Particles. J. Phys. (Moscow) Reiher, M. A Nonorthogonal State- 1945, 9, 449. Interaction Approach for Matrix Product State Wave Functions. J. Chem. Theory (309) Dancoff, S. M. Non-Adiabatic Meson Comput. 2016, 12, 5881–5894. Theory of Nuclear Forces. Phys. Rev. 1950, 78, 382–385.

70 (310) Hirata, S.; Head-Gordon, M. Time- corporating Dynamic and Static Corre- Dependent Density Functional Theory lation. ChemRxiv 2020, within the TammâĂŞDancoff Approxi- mation. Chem. Phys. Lett. 1999, 314, (318) Maitra, N. T.; Zhang, F.; Cave, R. J.; 291 – 299. Burke, K. Double Excitations within Time-Dependent Density Functional (311) Cordova, F.; Doriol, L. J.; Ipatov, A.; Theory Linear Response. J. Chem. Casida, M. E.; Filippi, C.; Vela, A. Trou- Phys. 2004, 120, 5932–5937. bleshooting Time-Dependent Density- Functional Theory for Photochemical (319) Elliott, P.; Goldson, S.; Canahui, C.; Applications: Oxirane. J. Chem. Phys. Maitra, N. T. Perspectives on Double- 2007, 127, 164111. Excitations in TDDFT. Chem. Phys. 2011, 391, 110 – 119. (312) Goerigk, L.; Casanova-Paéz, M. The Trip to the Density Functional Theory Zoo (320) Katriel, J.; Zahariev, F.; Burke, K. Sym- Continues: Making a Case for Time- metry and Degeneracy in Density Func- Dependent Double Hybrids for Excited- tional Theory. Int. J. Quantum Chem. State Problems. Aust. J. Chem. 2020, in 2001, 85, 432–435. press, DOI:10.1071/CH20093. (321) Shao, Y.; Head-Gordon, M.; Krylov, A. I. (313) Worth, G. A.; Cederbaum, L. S. Beyond The SpinâĂŞFlip Approach within Born-Oppenheimer: Molecular Dynam- Time-Dependent Density Functional ics Through a conical Intersection. Annu. Theory: Theory and Applications to Rev. Phys. Chem. 2004, 55, 127–158. Diradicals. J. Chem. Phys. 2003, 118, 4807–4818. (314) Doltsinis, N. L. Molecular Dynamics Be- yond the Born-Oppenheimer Approxi- (322) Gavnholt, J.; Olsen, T.; Engelund, M.; mation: Mixed Quantum-Classical Ap- Schiøtz, J. ∆ Self-Consistent Field proaches; NIC Series; John von Neuman Method to obtain Potential Energy Sur- Institut for computing, 2006; Vol. 31; pp faces of Excited Molecules on Surfaces. 389–409. Phys. Rev. B 2008, 78, 075441.

(315) Jacquemin, D.; Adamo, C. In Density- (323) Maurer, R. J.; Reuter, K. Assessing Functional Methods for Excited States; Computationally Efficient Isomerization Ferré, N., Filatov, M., Huix-Rotllant, M., Dynamics: ∆-SCF Density-Functional Eds.; Springer International Publishing: Theory Study of Azobenzene Molecular Cham, 2016; pp 347–375. Switching. J. Chem. Phys. 2011, 135, 224303. (316) Li, S. L.; Marenich, A. V.; Xu, X.; Truhlar, D. G. Configuration Interaction- (324) Maurer, R. J.; Reuter, K. Excited-State Corrected TammâĂŞDancoff Approxi- Potential-Energy Surfaces of Metal- mation: A Time-Dependent Density Adsorbed Organic Molecules from Lin- Functional Method with the Correct Di- ear Expansion ∆-Self-Consistent Field mensionality of Conical Intersections. J. Density-Functional Theory (∆SCF- Phys. Chem. Lett. 2014, 5, 322–328. DFT). J. Chem. Phys. 2013, 139, 014708. (317) Bannwarth, C.; Yu, J. K.; Hohen- stein, E. G.; Martínez, T. J. Hole- (325) Chai, J.-D.; Head-Gordon, M. System- Hole Tamm-Dancoff-Approximated Den- atic Optimization of Long-Range Cor- sity Functional Theory: A Highly Ef- rected Hybrid Density Functionals. J. ficient Electronic Structure Method In- Chem. Phys. 2008, 128, 084106.

71 (326) Tozer, D. J.; Handy, N. C. On the De- (334) Granucci, G.; Persico, M.; Spighi, G. termination of Excitation Energies using Surface Hopping Trajectory Simulations Density Functional Theory. Phys. Chem. with Spin-Orbit and Dynamical Cou- Chem. Phys. 2000, 2, 2117–2121. plings. J. Chem. Phys. 2012, 137, 22A501. (327) Ramakrishnan, R.; Hartmann, M.; Tapavicza, E.; von Lilienfeld, O. A. (335) Baer, M. Introduction to the Theory Electronic Spectra from TDDFT and of Electronic Non-Adiabatic Coupling Machine Learning in Chemical Space. J. Terms in Molecular Systems. Phys. Rep. Chem. Phys. 2015, 143, 084111. 2002, 358, 75–142.

(328) Dral, P. O.; Owens, A.; Dral, A.; (336) Kryachko, E. S.; Yarkony, D. R. Diabatic Csányi, G. Hierarchical Machine Learn- Bases and Molecular Properties. Int. J. ing of Potential Energy Surfaces. J. Quantum Chem. 2000, 76, 235–243. Chem. Phys. 2020, 152, 204110. (337) Marian, C. M. SpinâĂŞOrbit Coupling (329) Smith, J. S.; Nebgen, B. T.; Zu- and Intersystem Crossing in Molecules. batyuk, R.; Lubbers, N.; Devereuz, C.; WIREs Comput. Mol. Sci. 2012, 2, 187– Barros, K.; Tretiak, S.; Isayev, O.; Roit- 203. berg, A. E. Approaching Coupled Cluster Accuracy with a General-Purpose Neu- (338) K. G. Dyall, K. F. Introduction to ral Network Potential Through Transfer Relativistic Quantum Chemistry; Oxford Learning. Nat. Commun. 2019, 10 . University Press, 2007.

(330) Abedi, A.; Maitra, N. T.; Gross, E. (339) M. Reiher, A. W. Relativistic Quantum K. U. Exact Factorization of the Time- Chemistry; Wiley VCH Verlag Wein- Dependent Electron-Nuclear Wave Func- heim, 2009. tion. Phys. Rev. Lett. 2010, 105, 123002. (340) H. A. Bethe, E. E. S. Quantum Mechan- (331) Thachuk, M.; Ivanov, M. Y.; Ward- ics of One- and Two-Electron Atoms; law, D. M. A Semiclassical Approach Springer, Berlin, 1957. to IntenseâĂŘField AboveâĂŘThresh- (341) Pyykko, P. Relativistic Effects in Struc- old Dissociation in the Long Wavelength tural Chemistry. Chem. Rev. 1988, 88, Limit. J. Chem. Phys. 1996, 105, 4094– 563–594. 4104. (342) Neese, F.; Petrenko, T.; Ganyushin, D.; (332) Mitrić, R.; Petersen, J.; Bonači ć Olbrich, G. Advanced Aspects of Ab Koutecký, V. Laser-Field-Induced Initio Theoretical Optical Spectroscopy Surface-Hopping Method for the of Transition Metal complexes: Multi- Simulation and Control of Ultrafast plets, spin-orbit coupling and resonance Photodynamics. Phys. Rev. A 2009, 79, Raman intensities. Coord. Chem. Rev. 053416. 2007, 251, 288 – 327.

(333) Mitrić, R.; Petersen, J.; Wohlge- (343) Neese, F. Efficient and Accurate Ap- muth, M.; Werner, U.; Bonaçić- proximations to the Molecular Spin- Koutecký, V. Field-Induced Surface Orbit Coupling Operator and their Use Hopping Method for Probing Transition in Molecular G-Tensor Calculations. J. State Nonadiabatic Dynamics of Ag3. Chem. Phys. 2005, 122, 034107. Phys. Chem. Chem. Phys. 2011, 13, 8690–8696. (344) Richter, M.; Marquetand, P.; González- Vázquez, J.; Sola, I.; González, L.

72 SHARC: Ab Initio Molecular Dynamics (353) Wittenbrink, N.; Ndome, H.; Eisfeld, W. with Surface Hopping in the Adiabatic Toward SpinâĂŞOrbit coupled Diabatic Representation Including Arbitrary cou- Potential Energy Surfaces for Methyl Io- plings. J. Chem. Theory Comput. 2011, dide Using Effective Relativistic Cou- 7, 1253–1258. pling by Asymptotic Representation. J. Phys. Chem. A 2013, 117, 7408–7420. (345) Mai, S.; Marquetand, P.; González, L. A General Method to Describe Intersystem (354) Varga, Z.; Parker, K. A.; Truhlar, D. G. Crossing Dynamics in Trajectory Surface Direct Diabatization Based on Nonadi- Hopping. Int. J. Quantum Chem. 2015, abatic Couplings: The N/D Method. 115, 1215–1231. Phys. Chem. Chem. Phys. 2018, 20, 26643–26659. (346) Mai, S.; Plasser, F.; Marquetand, P.; GonzÃąlez, L. Attosecond Molecular Dy- (355) Nakamura, H.; Truhlar, D. G. Direct Di- namics; The Royal Society of Chemistry, abatization of Electronic States by the 2018; pp 348–385. Fourfold Way. II. Dynamical Correlation and Rearrangement Processes. J. Chem. (347) Köppel, H.; Gronki, J.; Mahapatra, S. Phys. 2002, 117, 5576–5593. Construction Scheme for Regularized Di- abatic States. J. Chem. Phys. 2001, 115, (356) Cave, R. J.; Stanton, J. F. Block Di- 2377–2388. agonalization of the Equation-of-Motion Coupled Cluster Effective Hamiltonian: (348) Richings, G. W.; Worth, G. A. A Prac- Treatment of Diabatic Potential Con- tical Diabatisation Scheme for Use with stants and Triple Excitations. J. Chem. the Direct-Dynamics Variational Multi- Phys. 2014, 140, 214112. configuration Gaussian Method. J. Phys. Chem. A 2015, 119, 12457–12470. (357) Venghaus, F.; Eisfeld, W. Block- Diagonalization as a Tool for the Robust (349) Accomasso, D.; Persico, M.; Granucci, G. Diabatization of High-Dimensional Po- Diabatization by Localization in the tential Energy Surfaces. J. Chem. Phys. Framework of Configuration Interaction 2016, 144, 114110. Based on Floating Occupation Molecular Orbitals (FOMO-CI). ChemPhotoChem (358) Robertson, C.; González-Vázquez, J.; 2019, 3, 933–944. corral, I.; Díaz-Tendero, S.; Díaz, C. Nonadiabatic Scattering of NO off Au3 (350) Lenzen, T.; Manthe, U. Neural Network Clusters: A Simple and Robust Dia- Based Coupled Diabatic Potential En- batic State Manifold Generation Method ergy Surfaces for Reactive Scattering. J. for Multiconfigurational Wavefunctions. Chem. Phys. 2017, 147, 084105. J. Comput. Chem. 2019, 40, 794–810. (351) Subotnik, J. E.; Yeganeh, S.; Cave, R. J.; (359) Li, J.; Jiang, B.; Guo, H. Permutation In- Ratner, M. A. Constructing Diabatic variant Polynomial Neural Network Ap- States from Adiabatic States: Extending proach to Fitting Potential Energy Sur- Generalized MullikenâĂŞHush to Multi- faces. II. Four-Atom Systems. J. Chem. ple Charge Centers with Boys Localiza- Phys. 2013, 139, 204103. tion. J. Chem. Phys. 2008, 129, 244101. (352) Hoyer, C. E.; Parker, K.; Gagliardi, L.; (360) Jiang, B.; Guo, H. Permutation Invariant Truhlar, D. G. The DQ and DQØ Elec- Polynomial Neural Network Approach tronic Structure Diabatization Methods: to Fitting Potential Energy Surfaces. J. Validation for General Applications. J. Chem. Phys. 2013, 139, 054112. Chem. Phys. 2016, 144, 194101.

73 (361) Jiang, B.; Guo, H. Permutation Invariant Marquetand, P.; González, L. Mixed Polynomial Neural Network Approach to Quantum-Classical Dynamics in the Fitting Potential Energy Surfaces. III. Adiabatic Representation to Simulate Molecule-Surface Interactions. J. Chem. Molecules Driven by Strong Laser Pulses. Phys. 2014, 141, 034109. J. Phys. Chem. A 2012, 116, 2800–2807.

(362) Jiang, B.; Li, J.; Guo, H. Potential En- (370) Köppel, H.; Domcke, W.; Ceder- ergy Surfaces from High Fidelity Fitting baum, L. S. Multimode Molecular Dy- of Ab Initio Points: The Permutation In- namics Beyond the Born-Oppenheimer variant Polynomial - Neural Network Ap- Approximation. Adv. Chem. Phys. 1984, proach. Int. Rev. Phys. Chem. 2016, 35, 57, 59–246. 479–506. (371) Ben-Nun, M.; Martínez, T. J. Advances (363) Mai, S.; M. Richter, M.; Marquetand, P.; in Chemical Physics; John Wiley & Sons, González, L. Excitation of Nucleobases Ltd, 2002; pp 439–512. from a Computational Perspective II: Dynamics. 2014, (372) Beck, M.; JÃďckle, A.; Worth, G.; Meyer, H.-D. The Multiconfiguration (364) Liu, W. Essentials of Relativistic Quan- Time-Dependent Hartree (MCTDH) tum Chemistry. J. Chem. Phys. 2020, Method: A Highly Efficient Algorithm 152, 180901. for Propagating Wavepackets. Phys. Rep. 2000, 324, 1–105. (365) Horton, S. L.; Liu, Y.; Forbes, R.; Makhija, V.; Lausten, R.; Stolow, A.; (373) Yeager, D. L.; Jørgensen, P. A Multicon- Hockett, P.; Marquetand, P.; Roz- figurational Time-Dependent Hartree- gonyi, T.; Weinacht, T. Excited state Fock Approach. Chem. Phys. Lett. 1979, dynamics of CH2I2 and CH2BrI studied 65, 77–80. with UV pump VUV probe photoelec- tron spectroscopy. J. Chem. Phys. 2019, (374) Manthe, U. Wavepacket Dynamics 150, 174201. and the Multi-Configurational Time- Dependent Hartree Approach. J. Phys.: (366) Horton, S. L.; Liu, Y.; Chakraborty, P.; Condens. Matter 2017, 29, 253001. Marquetand, P.; Rozgonyi, T.; Mat- sika, S.; Weinacht, T. Strong-Field- Ver- (375) Eng, J.; Gourlaouen, C.; Gin- sus Weak-Field-Ionization Pump-Probe densperger, E.; Daniel, C. Spin-Vibronic Spectroscopy. Phys. Rev. A 2018, 98, Quantum Dynamics for Ultrafast 053416. Excited-State Processes. Acc. Chem. Res. 2015, 48, 809–817. (367) Sussman, B. J.; Townsend, D.; Ivanov, M. Y.; Stolow, A. Dynamic (376) Gómez, S.; Heindl, M.; Szabadi, A.; Stark Control of Photochemical Pro- González, L. From Surface Hopping to cesses. Science 2006, 314, 278–281. Quantum Dynamics and Back. Find- ing Essential Electronic and Nuclear De- (368) Marquetand, P.; Richter, M.; González- grees of Freedom and Optimal Surface Vázquez, J.; Sola, I.; González, L. Nona- Hopping Parameters. J. Phys. Chem. A diabatic Ab Initio Molecular Dynamics 2019, 123, 8321–8332. Including Spin-Orbit Coupling and Laser Fields. Faraday Discuss. 2011, 153, 261– (377) Ischtwan, J.; Collins, M. A. Molecular 273. Potential Energy Surfaces by Interpola- tion. J. Chem. Phys. 1994, 100, 8080– (369) Bajo, J. J.; González-Vázquez, J.; 8088. Sola, I.; Santamaria, J.; Richter, M.;

74 (378) Evenhuis, C. R.; collins, M. A. Interpo- (386) Curchod, B. F. E.; Rauer, C.; Marque- lation of Diabatic Potential Energy Sur- tand, P.; González, L.; Martínez, T. J. faces. J. Chem. Phys. 2004, 121, 2515– Communication: GAIMSâĂŤGeneral- 2527. ized Ab Initio Multiple Spawning for Both Internal Conversion and Intersys- (379) Evenhuis, C.; Martínez, T. J. A Scheme tem Crossing Processes. J. Chem. Phys. to Interpolate Potential Energy Surfaces 2016, 144, 101102. and Derivative Coupling Vectors with- out Performing a Global Diabatization. (387) Mignolet, B.; Curchod, B. F. E. A Walk J. Chem. Phys. 2011, 135, 224110. Through the Approximations of Ab Ini- tio Multiple Spawning. J. Chem. Phys. (380) Mukherjee, S.; Bandyopadhyay, S.; 2018, 148, 134110. Paul, A. K.; Adhikari, S. Construction of Diabatic Hamiltonian Matrix from Ab (388) Freixas, V. M.; Fernandez-Alberti, S.; Initio Calculated Molecular Symmetry Makhov, D. V.; Tretiak, S.; Sha- Adapted Nonadiabatic Coupling Terms lashilin, D. An Ab Initio Multiple and Nuclear Dynamics for the Excited Cloning Approach for the Simulation of States of Na3 Cluster. J. Phys. Chem. A Photoinduced Dynamics in Conjugated 2013, 117, 3475–3495. Molecules. Phys. Chem. Chem. Phys. 2018, 20, 17762–17772. (381) Worth, G.; Robb, M.; Lasorne, B. Solv- ing the Time-Dependent Schrödinger (389) Tomislav Beguşić and Aurélien Patoz

Equation for Nuclear Motion in One and Miroslav Şulc and Jir, í Vaníçek, On- Step: Direct Dynamics of Non-Adiabatic the-Fly Ab Initio Three Thawed Gaus- Systems. Mol. Phys. 2008, 106, 2077– sians Approximation: A Semiclassical 2091. Approach to Herzberg-Teller Spectra. Chem. Phys. 2018, 515, 152 – 163. (382) Persico, M.; Granucci, G. An Overview of Nonadiabatic Dynamics Simulations (390) Markland, T.; Ceriotti, M. Nuclear Methods, with Focus on the Direct Ap- Quantum Effects Enter the Mainstream. proach Versus the Fitting of Poten- Nat. Rev. Chem. 2018, 2 . tial Energy Surfaces. Theor. Chem. Acc. (391) Miller, W. H. Classical S Matrix: Numer- 2014, 133, 1526. ical Application to Inelastic Collisions. J. (383) Komarova, K. G.; Remacle, F.; Chem. Phys. 1970, 53, 3578–3587. Levine, R. On the Fly Quantum (392) Ceotto, M.; Atahan, S.; Shim, S.; Dynamics of Electronic and Nuclear Tantardini, G. F.; Aspuru-Guzik, A. Wave Packets. Chem. Phys. Lett. 2018, First-Principles Semiclassical Initial 699, 155 – 161. Value Representation Molecular Dynam- (384) Lasorne, B.; Robb, M. A.; Worth, G. A. ics. Phys. Chem. Chem. Phys. 2009, 11, Direct Quantum Dynamics using Vari- 3861–3867. ational Multi-Configuration Gaussian (393) Nakamura, H.; Nanbu, S.; Teranishi, Y.; Wavepackets. Implementation Details Ohta, A. Development of Semiclassical and Test Case. Phys. Chem. Chem. Phys. Molecular Dynamics Simulation Method. 2007, 9, 3210–3227. Phys. Chem. Chem. Phys. 2016, 18, (385) Ben-Nun, M.; Martínez, T. J. Photo- 11972–11985. dynamics of Ethylene: Ab Initio Stud- (394) Gao, X.; Saller, M. A. C.; Liu, Y.; ies of Conical Intersections. Chem. Phys. Kelly, A.; Richardson, J. O.; Geva, E. 2000, 259, 237 – 248. Benchmarking Quasiclassical Mapping

75 Hamiltonian Methods for Simulating (403) Tully, J. C. Molecular Dynamics with Electronically Nonadiabatic Molecular Electronic Transitions. J. Chem. Phys. Dynamics. J. Chem. Theory Comput. 1990, 93, 1061–1071. 2020, 16, 2883–2895. (404) Tully, J. C. Nonadiabatic Molecular Dy- (395) Ceriotti, M.; More, J.; Manolopou- namics. Int. J. Quantum Chem. 1991, los, D. E. i-PI: A Python Interface for Ab 40, 299–309. Initio Path Integral Molecular Dynamics Simulations. Comput. Phys. Commun. (405) C. Tully, J. Mixed QuantumâĂŞClassical 2014, 185, 1019 – 1026. Dynamics. Faraday Discuss. 1998, 110, 407–419. (396) Kapil, V. et al. i-PI 2.0: A Univer- sal Force Engine for Advanced Molec- (406) S. Mai, L. G., P. Marquetand In Quan- ular Simulations. Comput. Phys. Com- tum Chemistry and Dynamics of Ex- mun. 2019, 236, 214–223. cited States: Methods and Applications; González, L., Lindh, R., Eds.; Wiley, (397) Thoss, M.; Miller, W. H.; Stock, G. 2020; in press. Semiclassical Description of Nonadia- batic Quantum Dynamics: Application (407) Zener, C. Non-Adiabatic Crossing of En- to the S1âĂŞS2 Conical Intersection in ergy Levels. Proc. Roy. Soc. Lond. A Pyrazine. J. Chem. Phys. 2000, 112, 1932, 137, 696–701. 10282–10292. (408) Wittig, C. The Landau-Zener Formula. (398) Lee, M. K.; Huo, P.; Coker, D. F. Semi- J. Phys. Chem. B 2005, 109, 8428–8430. classical Path Integral Dynamics: Pho- (409) Zhu, C.; Kamisaka, H.; Nakamura, H. tosynthetic Energy Transfer with Re- Significant Improvement of the Tra- alistic Environment Interactions. Annu. jectory Surface Hopping Method by Rev. Phys. Chem. 2016, 67, 639–668. the ZhuâĂŞNakamura Theory. J. Chem. (399) Stock, G.; Thoss, M. Semiclassical De- Phys 2001, 115, 11036–11039. scription of Nonadiabatic Quantum Dy- (410) Zhu, C.; Kamisaka, H.; Nakamura, H. namics. Phys. Rev. Lett. 1997, 78, 578– New Implementation of the Trajectory 581. Surface Hopping Method with Use of the (400) Westermayr, J.; Marquetand, P. Ma- ZhuâĂŞNakamura Theory. II. Applica- chine Learning and Excited-State Molec- tion to the Charge Transfer Processes in ular Dynamics. Mach. Learn.: Sci. Tech- the 3D DH2+ System. J. Chem. Phys. nol. 2020, in press, doi:10.1088/2632– 2002, 116, 3234–3247. 2153/ab9c3e. (411) Oloyede, P.; MilâĂŹnikov, G.; Naka- (401) Weinreich, J.; Römer, A.; Pale- mura, H. Generalized Trajectory Sur- ico, M. L.; Behler, J. Properties of face Hopping Method Based on the Zhu- α-Brass Nanoparticles. 1. Neural Net- Nakamura Theory. J. Chem. Phys. 2006, work Potential Energy Surface. J. Phys. 124, 144110. Chem. C 2020, in press. (412) Ishida, T.; Nanbu, S.; Nakamura, H. (402) Lin, Q.; Zhang, Y.; Zhao, B.; Jiang, B. Clarification of Nonadiabatic Chemical Automatically Growing Global Reactive Dynamics by the Zhu-Nakamura Theory Neural Network Potential Energy Sur- of Nonadiabatic Transition: From Tri- faces: A Trajectory-Free Active Learn- Atomic Systems to Reactions in Solu- ing Strategy. J. Chem. Phys. 2020, 152, tions. Int. Rev. Phys. Chem. 2017, 36, 154104. 229–286.

76 (413) Zhu, L.; Kleiman, V.; Li, X.; Lu, S. P.; (423) Mai, S.; Richter, M.; Marquetand, P.; Trentelman, K.; Gordon, R. J. Ultrafast González, L. The DNA Nucleobase coherent control and Destruction of Exci- Thymine in Motion – Intersystem Cross- tons in Quantum Wells. Phys. Rev. Lett. ing Simulated with Surface Hopping. 1995, 75, 2598–2601. Chem. Phys. 2017, 482, 9 – 15.

(414) Granucci, G.; Persico, M. Critical Ap- (424) Raghunathan Ramakrish- praisal of the Fewest Switching Algo- nan, M. R., Pavlo O. Dral; von rithm for Surface Hopping. J. Chem. Lilienfeld, O. A. Quantum Chem- Phys. 2007, 126, 134114. istry Structures and Properties of 134 Kilo Molecules. Sci. Data 2014, 1 . (415) Fabiano, E.; Keal, T.; Thiel, W. Imple- mentation of Surface Hopping Molecular (425) Artrith, N.; Morawietz, T.; Behler, J. Dynamics using Semiempirical Methods. High-Dimensional Neural-Network Po- Chem. Phys. 2008, 349, 334 – 347. tentials for Multicomponent Systems: Applications to Zinc Oxide. Phys. Rev. (416) Malhado, J. P.; Bearpark, M. J.; B 2011, 83, 153101. Hynes, J. T. Non-Adiabatic Dynam- ics Close to Conical Intersections and (426) Huang, B.; von Lilienfeld, O. A. Commu- the Surface Hopping Perspective. Front. nication: Understanding Molecular Rep- Chem. 2014, 2, 97. resentations in Machine Learning: The Role of Uniqueness and Target Similar- (417) Wang, L.; Akimov, A.; Prezhdo, O. V. ity. J. Chem. Phys. 2016, 145, 161102. Recent Progress in Surface Hopping: 2011-2015. J. Phys. Chem. Lett. 2016, (427) Yao, K.; Herr, J. E.; Toth, D. W.; Mck- 7, 2100–2112. intyre, R.; Parkhill, J. The TensorMol- 0.1 Model Chemistry: A Neural Network (418) Subotnik, J. E.; Rhee, Y. M. On Surface Augmented with Long-Range Physics. Hopping and Time-Reversal. J. Phys. Chem. Sci. 2018, 9, 2261–2269. Chem. A 2015, 119, 990–995. (428) Schütt, K. T.; Sauceda, H. E.; Kinder- (419) HammesâĂŘSchiffer, S.; Tully, J. C. Pro- mans, P.-J.; Tkatchenko, A.; Müller, K.- ton Transfer in Solution: Molecular Dy- R. SchNet – A Deep Learning Archi- namics with Quantum Transitions. J. tecture for Molecules and Materials. J. Chem. Phys. 1994, 101, 4657–4667. Chem. Phys. 2018, 148, 241722.

(420) Sawada, S.-I.; Nitzan, A.; Metiu, H. (429) Nebgen, B.; Lubbers, N.; Smith, J. S.; Mean-Trajectory Approximation for Sifain, A. E.; Lokhov, A.; Isayev, O.; Charge- and Energy-Transfer Processes Roitberg, A. E.; Barros, K.; Tretiak, S. at Surfaces. Phys. Rev. B 1985, 32, Transferable Dynamic Molecular Charge 851–867. Assignment Using Deep Neural Net- (421) Li, X.; Tully, J. C.; Schlegel, H. B.; works. J. Chem. Theory Comput. 2018, Frisch, M. J. Ab Initio Ehrenfest Dynam- 14, 4687–4698. ics. J. Chem. Phys. 2005, 123, 084106. (430) Sifain, A. E.; Lubbers, N.; Nebgen, B. T.; (422) Mai, S.; Marquetand, P.; González, L. Smith, J. S.; Lokhov, A. Y.; Isayev, O.; Intersystem Crossing Pathways in the Roitberg, A. E.; Barros, K.; Tretiak, S. Noncanonical Nucleobase 2-Thiouracil: Discovering a Transferable Charge As- A Time-Dependent Picture. J. Phys. signment Model Using Machine Learn- Chem. Lett. 2016, 7, 1978–1983. ing. J. Phys. Chem. Lett. 2018, 9, 4495– 4501.

77 (431) Schütt, K. T.; Gastegger, M.; Simulating Light-Induced Processes in Tkatchenko, A.; Müller, K.-R. Ex- DNA. Molecules 2016, 22, 49. plainable AI: Interpreting, Explaining and Visualizing Deep Learning; Springer (440) Nogueira, J. J.; González, L. Computa- International Publishing, 2019; pp tional Photophysics in the Presence of an 311–330. Environment. Annu. Rev. Phys. Chem. 2018, 69, 473–497. (432) Schütt, K. T.; Kessel, P.; Gastegger, M.; Nicoli, K. A.; Tkatchenko, A.; (441) Barbatti, M.; Granucci, G.; Persico, M.; Müller, K.-R. SchNetPack: A Deep Ruckenbauer, M.; Vazdar, M.; Eckert- Learning Toolbox For Atomistic Sys- Maksić, M.; Lischka, H. The on-the- tems. J. Chem. Theory Comput. 2019, Fly Surface-Hopping Program System 15, 448–455. Newton-X: Application to Ab Initio Sim- ulation of the Nonadiabatic Photody- (433) Christensen, A. S.; Faber, F. A.; von namics of Benchmark Systems. J. Pho- Lilienfeld, O. A. Operators in Quantum tochem. Photobiol. A 2007, 190, 228– Machine Learning: Response Properties 240. in Chemical Space. J. Chem. Phys. 2019, 150, 064105. (442) Dral, P. O. Quantum Chemistry in the Age of Machine Learning. J. Phys. (434) Veit, M.; Wilkins, D. M.; Yang, Y.; Chem. Lett. 2020, 11, 2336–2347. Jr., R. A. D.; Ceriotti, M. Predicting Molecular Dipole Moments by Combin- (443) Chen, W.-K.; Fang, W.-H.; Cui, G. Inte- ing Atomic Partial Charges and Atomic grating Machine Learning with the Mul- Dipoles. arXiv 2020, 2003.12437 . tilayer Energy-Based Fragment Method for Excited States of Large Systems. J. (435) Gastegger, M.; Marquetand, P. Molecu- Phys. Chem. Lett. 2019, 10, 7836–7841. lar dynamics with neural-network poten- tials. arXiv:1812.07676 [physics.chem- (444) Chen, W.-K.; Zhang, Y.; Jiang, B.; ph] 2018, Fang, W.-H.; Cui, G. Efficient Con- struction of Excited-State Hessian Ma- (436) Thomas, M.; Brehm, M.; Fligg, R.; trices with Machine Learning Accel- Vöhringer, P.; Kirchner, B. Comput- erated Multilayer Energy-Based Frag- ing Vibrational Spectra from Ab Ini- ment Method. The Journal of Phys- tio Molecular Dynamics. Phys. Chem. ical Chemistry A 2020, in press, Chem. Phys. 2013, 15, 6608–6622. DOI:10.1021/acs.jpca.0c04117.

(437) Wilke, J.; Wilke, M.; Meerts, W. L.; (445) Behler, J.; Martoák, R.; Donadio, D.; Schmitt, M. Determination of Ground Parrinello, M. Metadynamics Simula- and Excited State Dipole Moments tions of the High-Pressure Phases of Sili- via Electronic Stark Spectroscopy: 5- con Employing a High-Dimensional Neu- Methoxyindole. J. Chem. Phys. 2016, ral Network Potential. Phys. Rev. Lett. 144, 044201. 2008, 100, 185501.

(438) Tennyson, J. Perspective: Accurate (446) Gastegger, M.; Schwiedrzik, L.; Bit- Ro-Vibrational Calculations on Small termann, M.; Berzsenyi, F.; Marque- Molecules. J. Chem. Phys. 2016, 145, tand, P. wACSF – Weighted Atom- 120901. Centered Symmetry Functions as De- scriptors in Machine Learning Potentials. (439) Marquetand, P.; Nogueira, J.; Mai, S.; J. Chem. Phys. 2018, 148, 241709. Plasser, F.; González, L. Challenges in

78 (447) Chen, M. S.; Zuehlsdorff, T. J.; Moraw- Bowman, J. M.; Truhlar, D. G. Direct ietz, T.; Isborn, C. M.; Markland, T. E. Diabatization and Analytic Representa- Exploiting Machine Learning to Effi- tion of Coupled Potential Energy Sur- ciently Predict Multidimensional Optical faces and Couplings for the Reactive Spectra in Complex Environments. arXiv Quenching of the Excited 2Σ+ State of 2020, 2005.09776 . OH by Molecular Hydrogen. J. Chem. Phys. 2019, 151, 104311. (448) Koch, W.; Zhang, D. H. Communica- tion: Separable Potential Energy Sur- (456) Yarkony, D. R. On the Consequences of faces from Multiplicative Artificial Neu- Nonremovable Derivative Couplings. I. ral Networks. J. Chem. Phys. 2014, 141, The Geometric Phase and Quasidiabatic 021101. States: A Numerical Study. J. Chem. Phys. 1996, 105, 10456–10461. (449) He, D.; Yuan, J.; Li, H.; Chen, M. Global Diabatic Potential Energy Surfaces and (457) Yarkony, D. R. On the Role of Coni- Quantum Dynamical Studies for the cal Intersections in Photodissociation. V. 1 + 1 + Li(2p) + H2(X Σg ) → LiH(X Σ ) + H Conical Intersections and the Geometric Reaction. Sci. Rep. 2016, 6 . Phase in the Photodissociation of Methyl Mercaptan. J. Chem. Phys. 1996, 104, (450) Guan, Y.; Fu, B.; Zhang, D. H. Con- 7866–7881. struction of Diabatic Energy Surfaces for LiFH with Artificial Neural Networks. J. (458) Ryabinkin, I. G.; Joubert-Doriol, L.; Iz- Chem. Phys. 2017, 147, 224307. maylov, A. F. When Do We Need to Ac- count for the Geometric Phase in Excited (451) Wang, S.; Yang, Z.; Yuan, J.; Chen, M. State Dynamics? J. Chem. Phys. 2014, New Diabatic Potential Energy Surfaces 140, 214116. of the NaH2 System and Dynamics Stud- ies for the Na(3p) + H2 → NaH + H (459) Gherib, R.; Ryabinkin, I. G.; Iz- Reaction. Sci. Rep. 2018, 8 . maylov, A. F. Why Do Mixed Quantum- Classical Methods Describe Short-Time (452) Yuan, J.; He, D.; Wang, S.; Chen, M.; Dynamics through Conical Intersections Han, K. Diabatic Potential Energy Sur- So Well? Analysis of Geometric Phase faces of MgH+ and Dynamic Studies for 2 Effects. J. Chem. Theory Comput. 2015, the Mg+(3p) +H → MgH+ + H Re- 2 11, 1375–1382. action. Phys. Chem. Chem. Phys. 2018, 20, 6638–6647. (460) Ryabinkin, I. G.; Joubert-Doriol, L.; Iz- maylov, A. F. Geometric Phase Effects (453) Yin, Z.; Guan, Y.; Fu, B.; Zhang, D. H. in Nonadiabatic Dynamics Near Conical Two-State Diabatic Potential Energy Intersections. Acc. Chem. Res. 2017, 50, Surfaces of ClH Based on Nonadiabatic 2 1785–1793. Couplings with Neural Networks. Phys. Chem. Chem. Phys. 2019, 21, 20372– (461) Plasser, F.; Ruckenbauer, M.; Mai, S.; 20383. Oppel, M.; Marquetand, P.; González, L. Efficient and Flexible Computation of (454) Akimov, A. V. A Simple Phase correc- Many-Electron Wave Function Overlaps. tion Makes a Big Difference in Nona- J. Chem. Theory Comput. 2016, 12, diabatic Molecular Dynamics. J. Phys. 1207. Chem. Lett. 2018, 9, 6096–6102. (455) Shu, Y.; Kryven, J.; Sampaio de Oliveira- (462) Zhu, X.; Yarkony, D. R. Toward Elim- Filho, A. G.; Zhang, L.; Song, G.-L.; inating the Electronic Structure Bottle- Li, S. L.; Meana-Pan˜eda, R.; Fu, B.; neck in Nonadiabatic Dynamics on the

79 Fly: An Algorithm to Fit Nonlocal, (469) Tapavicza, E.; Tavernelli, I.; Rothlis- Quasidiabatic, Coupled Electronic State berger, U. Trajectory Surface Hopping Hamiltonians Based on Ab Initio Elec- within Linear Response Time-Dependent tronic Structure Data. J. Chem. Phys. Density-Functional Theory. Phys. Rev. 2010, 132, 104101. Lett. 2007, 98, 023001.

(463) Zhu, X.; Yarkony, D. R. Quasi-Diabatic (470) Tavernelli, I.; Tapavicza, E.; Rothlis- Representations of Adiabatic Potential berger, U. Nonadiabatic Coupling Vec- Energy Surfaces Coupled by Conical In- tors within Linear Response Time- tersections including Bond Breaking: A Dependent Density Functional Theory. J. More General Construction Procedure Chem. Phys. 2009, 130, 124107. and an Analysis of the Diabatic Rep- resentation. J. Chem. Phys. 2012, 137, (471) Tavernelli, I.; Tapavicza, E.; Roth- 22A511. lisberger, U. Non-Adiabatic Dynamics using Time-Dependent Density Func- (464) Zhu, X.; Yarkony, D. R. On the Repre- tional Theory: Assessing the Coupling sentation of Coupled Adiabatic Potential Strengths. J. Mol. Struct.: THEOCHEM Energy Surfaces using Quasi-Diabatic 2009, 914, 22 – 29. Hamiltonians: A Distributed Origins Ex- pansion Approach. J. Chem. Phys. 2012, (472) Barbatti, M.; Aquino, A. J. A.; Lis- 136, 174110. chka, H. Ultrafast Two-Step Process in the Non-Adiabatic Relaxation of the (465) Rasmussen, C. E. In Advanced Lec- CH2NH2 Molecule. Mol. Phys. 2006, tures on Machine Learning: ML Sum- 104, 1053–1060. mer Schools 2003, Canberra, Australia, February 2 - 14, 2003, Tübingen, Ger- (473) Tao, H.; Allison, T. K.; Wright, T. W.; many, August 4 - 16, 2003, Revised Lec- Stooke, A. M.; Khurmi, C.; van tures; Bousquet, O., von Luxburg, U., Tilborg, J.; Liu, Y.; Falcone, R. W.; Rätsch, G., Eds.; Springer Berlin Heidel- Belkacem, A.; Martinez, T. J. Ultrafast berg: Berlin, Heidelberg, 2004; pp 63–71. internal conversion in ethylene. I. The excited state lifetime. J. Chem. Phys. (466) Mai, S.; Richter, M.; Ruckenbauer, M.; 2011, 134, 244306. Oppel, M.; Marquetand, P.; González, L. SHARC2.0: Surface Hopping Including (474) Allison, T. K.; Tao, H.; Glover, W. J.; ARbitrary Couplings – Program Pack- Wright, T. W.; Stooke, A. M.; age for Non-Adiabatic Dynamics. sharc- Khurmi, C.; van Tilborg, J.; Liu, Y.; md.org, 2018. Falcone, R. W.; Martínez, T. J.; Belka- cem, A. Ultrafast internal conversion in (467) Beard, E. J.; Sivaraman, G.; Vázquez- ethylene. II. Mechanisms and pathways Mayagoitia, A.; Vishwanath, V.; for quenching and hydrogen elimination. Cole, J. M. Comparative Dataset of J. Chem. Phys. 2012, 136, 124317. Experimental and Computational At- tributes of UV/Vis Absorption Spectra. (475) Mori, T.; Glover, W. J.; Schuur- Sci. Data 2019, 6 . man, M. S.; Martinez, T. J. Role of Ry- dberg States in the Photochemical Dy- (468) Barbatti, M.; Ruckenbauer, M.; Lis- namics of Ethylene. J. Phys. Chem. A chka, H. The Photodynamics of Ethy- 2012, 116, 2808–2818. lene: A Surface-Hopping Study on Struc- tural Aspects. J. Chem. Phys. 2005, 122, (476) Sellner, B.; Barbatti, M.; Müller, T.; 174307. Domcke, W.; Lischka, H. Ultrafast Non- Adiabatic Dynamics of Ethylene includ-

80 ing Rydberg States. Mol. Phys. 2013, (485) Hansen, K.; Montavon, G.; Biegler, F.; 111, 2439–2450. Fazli, S.; Rupp, M.; Scheffler, M.; von Lilienfeld, O. A.; Tkatchenko, A.; (477) Barbatti, M.; Lan, Z.; Crespo-Otero, R.; Müller, K.-R. Assessment and Valida- Szymczak, J. J.; Lischka, H.; Thiel, W. tion of Machine Learning Methods for Critical Appraisal of Excited State Nona- Predicting Molecular Atomization Ener- diabatic Dynamics Simulations of 9H- gies. J. Chem. Theory Comput. 2013, 9, Adenine. J. Chem. Phys. 2012, 117, 3404–3419. 22A503. (486) Chmiela, S.; Tkatchenko, A.; (478) Hollas, D.; Šištík, L.; Hohenstein, E. G.; Sauceda, H. E.; Poltavsky, I.; Martínez, T. J.; Slavíček, P. Nonadi- Schütt, K. T.; Müller, K.-R. Machine abatic Ab Initio Molecular Dynamics Learning of Accurate Energy-Conserving with the Floating Occupation Molecu- Molecular Force Fields. Sci. Adv. 2017, lar Orbital-Complete Active Space Con- 3 . figuration Interaction Method. J. Chem. Theory Comput. 2018, 14, 339–350. (487) Christensen, A. S.; Bratholm, L. A.; Faber, F. A.; Anatole von Lilienfeld, O. (479) Botu, V.; Ramprasad, R. Adaptive Ma- FCHL Revisited: Faster and More Ac- chine Learning Framework to Accelerate curate Quantum Machine Learning. J. Ab Initio Molecular Dynamics. Int. J. Chem. Phys. 2020, 152, 044107. Quant. Chem. 2015, 115, 1074–1083. (488) Kim, H.; Park, J.; Choi, S. Energy Re- (480) Behler, J. Constructing High- finement and Analysis of Structures in Dimensional Neural Network Potentials: the QM9 Database via a Highly Accurate A Tutorial Review. Int. J. Quantum Quantum Chemical Method. Sci. Data Chem. 2015, 115, 1032–1050. 2019, 6 . (481) Ceriotti, M.; Tribello, G. A.; Par- (489) Glavatskikh, M.; Leguy, J.; Hunault, G.; rinello, M. Demonstrating the Trans- Cauchi, T.; Da Mota, B. Dataset’s ferability and the Descriptive Power of Chemical Diversity Limits the Generaliz- Sketch-Map. J. Chem. Theory Comput. ability of Machine Learning Predictions. 2013, 9, 1521–1532. J. Cheminform. 2019, 11 . (482) Dral, P. O.; Owens, A.; Yurchenko, S. N.; (490) https://www.kaggle.com/c/champs- Thiel, W. Structure-Based Sampling and scalar-coupling/, 2020-05-01. Self-Correcting Machine Learning for Ac- curate Calculations of Potential En- (491) von Lilienfeld, O. A. The QM9 challenge. ergy Surfaces and Vibrational Levels. J. https://twitter.com/ProfvLilienfeld/status/1073179005854121984, Chem. Phys. 2017, 146, 244108. 2018.

(483) Sobol’, I. M.; Asotsky, D.; Kreinin, A.; (492) Fink, T.; Bruggesser, H.; Reymond, J.- Kucherenko, S. Construction and Com- L. Virtual Exploration of the Small- parison of High-Dimensional Sobol’ Gen- Molecule Chemical Universe Below 160 erators. Wilmott 2011, 2011, 64–79. Daltons. Angew. Chem., Int. Ed. 2005, 44, 1504–1508. (484) Uteva, E.; Graham, R. S.; Wilkin- son, R. D.; Wheatley, R. J. Active Learn- (493) Fink, T.; Reymond, J.-L. Virtual Ex- ing in Gaussian Process Interpolation ploration of the Chemical Universe up of Potential Energy Surfaces. J. Chem. to 11 Atoms of C, N, O, F: Assembly Phys. 2018, 149, 174114. of 26.4 Million Structures (110.9 Mil- lion Stereoisomers) and Analysis for New

81 Ring Systems, Stereochemistry, Physico- (501) Perdew, J. P.; Burke, K.; Ernzer- chemical Properties, Compound Classes, hof, M. Generalized Gradient Approxi- and Drug Discovery. J. Chem. Inf. mation Made Simple. Phys. Rev. Lett. Model. 2007, 47, 342–353. 1996, 77, 3865–3868.

(494) Blum, L. C.; Reymond, J.-L. 970 Mil- (502) Nakata, M.; Shimazaki, T. PubChemQC lion Druglike Small Molecules for Vir- Project: A Large-Scale First-Principles tual Screening in the Chemical Universe Electronic Structure Database for Data- Database GDB-13. J. Am. Chem. Soc. Driven Chemistry. J. Chem. Inf. Model. 2009, 131, 8732. 2017, 57, 1300–1308.

(495) Montavon, G.; Rupp, M.; Gobre, V.; (503) Kolb, B.; Zhao, B.; Li, J.; Jiang, B.; Vazquez-Mayagoitia, A.; Hansen, K.; Guo, H. Permutation Invariant Potential Tkatchenko, A.; Müller, K.-R.; von Energy Surfaces for Polyatomic Reac- Lilienfeld, O. A. Machine Learning tions using Atomistic Neural Networks. of Molecular Electronic Properties in J. Chem. Phys. 2016, 144, 224103. Chemical Compound Space. New J. Phys. 2013, 15, 095003. (504) Bernstein, N.; Csányi, G.; Deringer, V. L. De Novo Exploration and Self-Guided (496) Ruddigkeit, L.; van Deursen, R.; Learning of Potential-Energy Surfaces. Blum, L. C.; Reymond, J.-L. Enu- npj Comput. Mater. 2019, 5 . meration of 166 Billion Organic Small Molecules in the Chemical Universe (505) Yao, Z.; Sanchez-Lengeling, B.; Bob- Database GDB-17. J. Chem. Inf. Model. bitt, N. S.; Bucior, B. J.; Kumar, S. 2012, 52, 2864–2875. G. H.; Collins, S. P.; Burns, T.; Woo, T. K.; Farha, O.; Snurr, R. Q.; (497) Ramakrishnan, R.; Dral, P. O.; Aspuru-Guzik, A. Inverse Design Rupp, M.; von Lilienfeld, O. A. Big of Nanoporous Crystalline Retic- Data Meets Quantum Chemistry Ap- ular Materials with Deep Gen- proximations: The ∆-Machine Learning erative Models. ChemRxiv 2020, Approach. J. Chem. Theory Comput. DOI:10.26434/chemrxiv.12186681.v1 . 2015, 11, 2087–2096. (506) Krenn, M.; Häse, F.; Nigam, A.; (498) Lee, C.; Yang, W.; Parr, R. G. Devel- Friederich, P.; Aspuru-Guzik, A. SELF- opment of the Colle-Salvetti Correlation- IES: A Robust Representation of Seman- Energy Formula into a Functional of the tically Constrained Graphs with an Ex- Electron Density. Phys. Rev. B 1988, 37, ample Application in Chemistry. arXiv 785–789. 2019, abs/1905.13741 .

(499) Becke, A. D. Density-Functional (507) Gebauer, N. W. A.; Gastegger, M.; Exchange-Energy Approximation with Schütt, K. T. Symmetry-Adapted Gen- Correct Asymptotic Behavior. Phys. eration of 3D Point Sets for the Targeted Rev. A 1988, 38, 3098–3100. Discovery of Molecules. arXiv 2019, 1906.00957 . (500) Stuke, A.; Kunkel, C.; Golze, D.; Todor- ović, M.; Margraf, J. T.; Reuter, K.; (508) Schütt, K. T.; Arbabzadah, F.; Rinke, P.; Oberhofer, H. Atomic Struc- Chmiela, S.; Müller, K. R.; tures and Orbital Energies of 61,489 Tkatchenko, A. Quantum-Chemical Crystal-Forming Organic Molecules. Sci. Insights from Deep Tensor Neural Net- Data 2020, 7 . works. Nat. Commun. 2017, 8, 13890 EP –.

82 (509) Ye, S.; Hu, W.; Li, X.; Zhang, J.; Moving Least-Squares Methods for Fit- Zhong, K.; Zhang, G.; Luo, Y.; ting Potential Energy Surfaces: Comput- Mukamel, S.; Jiang, J. A Neural Network ing High-Density Potential Energy Sur- Protocol for Electronic Excitations of N- face Data from Low-Density Ab Initio Methylacetamide. Proc. Natl. Acad. Sci. Data Points. J. Chem. Phys. 2007, 126, 2019, 116, 11612–11617. 184108. (510) Jorgensen, W. L.; Maxwell, D. S.; (518) Dawes, R.; Thompson, D. L.; Wag- Tirado-Rives, J. Development and Test- ner, A. F.; Minkoff, M. Interpolating ing of the OPLS All-Atom Force Field Moving Least-Squares Methods for Fit- on Conformational Energetics and Prop- ting Potential Energy Surfaces: A Strat- erties of Organic Liquids. J. Am. Chem. egy for Efficient Automatic Data Point Soc. 1996, 118, 11225–11236. Placement in High Dimensions. J. Chem. Phys. 2008, 128, 084107. (511) Van Der Spoel, D.; Lindahl, E.; Hess, B.; Groenhof, G.; Mark, A. E.; Berend- (519) Braams, B. J.; Bowman, J. M. Permuta- sen, H. J. C. GROMACS: Fast, Flexi- tionally Invariant Potential Energy Sur- ble, and Free. J. Comp. Chem. 2005, 26, faces in High Dimensionality. Int. Rev. 1701–1718. Phys. Chem. 2009, 28, 577–606. (512) Ceriotti, M.; Tribello, G. A.; Par- (520) Qu, C.; Yu, Q.; Bowman, J. M. Permuta- rinello, M. Simplifying the Representa- tionally Invariant Potential Energy Sur- tion of Complex Free-Energy Landscapes faces. Annu. Rev. Phys. Chem. 2018, 69, using Sketch-Map. Proc. Natl. Acad. Sci 151–175. 2011, 108, 13023–13028. (521) Lorenz, S.; Groç, A.; Scheffler, M. Rep- (513) Tribello, G. A.; Ceriotti, M.; Par- resenting High-Dimensional Potential- rinello, M. Using Sketch-Map Coordi- Energy Surfaces for Reactions at Surfaces nates to Analyze and Bias Molecular Dy- by Neural Networks. Chem. Phys. Lett. namics Simulations. Proc. Natl. Acad. 2004, 395, 210 – 215. Sci 2012, 109, 5196–5201. (522) Raff, L. M.; Malshe, M.; Hagan, M.; (514) Seung, H. S.; Opper, M.; Sompolinsky, H. Doughan, D. I.; Rockley, M. G.; Koman- Query by Committee. Proceedings of the duri, R. Ab Initio Potential-Energy Sur- Fifth Annual Workshop on Computa- faces for Complex, Multichannel Systems tional Learning Theory. New York, NY, using Modified Novelty Sampling and USA, 1992; p 287âĂŞ294. Feedforward Neural Networks. J. Chem. Phys. 2005, 122, 084104. (515) Collins, M. Molecular Potential-Energy Surfaces for Chemical Reaction Dy- (523) Behler, J.; Parrinello, M. Generalized namics. Theor. Chem. Acc. 2002, 108, Neural-Network Representation of High- 313âĂŞ324. Dimensional Potential-Energy Surfaces. Phys. Rev. Lett. 2007, 98, 146401. (516) Godsi, O.; Collins, M. A.; Peskin, U. Quantum GrowâĂŤA Quantum Dynam- (524) Chen, J.; Xu, X.; Xu, X.; Zhang, D. H. A ics Sampling Approach for Growing Po- Global Potential Energy Surface for the tential Energy Surfaces and Nonadia- H2 + OH ↔ H2O + H Reaction using batic Couplings. J. Chem. Phys. 2010, Neural Networks. J. Chem. Phys. 2013, 132, 124106. 138, 154301. (517) Dawes, R.; Thompson, D. L.; Guo, Y.; (525) Jiang, B.; Guo, H. Dynamics of Water Wagner, A. F.; Minkoff, M. Interpolating Dissociative Chemisorption on Ni(111):

83 Effects of Impact Sites and Incident An- Efficient and Accurate Machine Learn- gles. Phys. Rev. Lett. 2015, 114, 166101. ing with a Physically Inspired Represen- tation. J. Phys. Chem. Lett. 2019, 10, (526) Shen, X.; Chen, J.; Zhang, Z.; Shao, K.; 4962–4967. Zhang, D. H. Methane Dissociation on Ni(111): A Fifteen-Dimensional Poten- (534) Wigner, E. On The Quantum Correction tial Energy Surface using Neural Net- for Thermodynamic Equilibrium. Phys. work Method. J. Chem. Phys. 2015, 143, Rev. 1932, 40, 749–750. 144701. (535) Bruccoleri, R. E.; Karplus, M. Conforma- (527) Shao, K.; Chen, J.; Zhao, Z.; tional Sampling using High-Temperature Zhang, D. H. Communication: Fit- Molecular Dynamics. Biopolymers 1990, ting Potential Energy Surfaces with 29, 1847–1862. Fundamental Invariant Neural Network. J. Chem. Phys. 2016, 145, 071101. (536) Maximova, T.; Moffatt, R.; Ma, B.; Nussinov, R.; Shehu, A. Principles and (528) Cui, J.; Krems, R. V. Efficient Non- Overview of Sampling Methods for Mod- Parametric Fitting of Potential Energy eling Macromolecular Structure and Dy- Surfaces for Polyatomic Molecules with namics. PLOS computational Biology Gaussian Processes. J. Phys. B: At., Mol. 2016, 12, 1–70. Opt. Phys. 2016, 49, 224001. (537) Kästner, J. Umbrella Sampling. Wi- (529) Kolb, B.; Marshall, P.; Zhao, B.; ley Interdiscip. Rev. Comput. Mol. Sci. Jiang, B.; Guo, H. Representing Global 2011, 1, 932–942. Reactive Potential Energy Surfaces Us- ing Gaussian Processes. J. Phys. Chem. (538) Tao, G. Trajectory-Guided Sampling for A 2017, 121, 2552–2557. Molecular Dynamics Simulation. Theor. Chem. Acc. 2019, 138, 34. (530) Kolb, B.; Luo, X.; Zhou, X.; Jiang, B.; Guo, H. High-Dimensional Atom- (539) Yang, Y. I.; Shao, Q.; Zhang, J.; istic Neural Network Potentials for Yang, L.; Gao, Y. Q. Enhanced Sampling MoleculeâĂŞSurface Interactions: HCl in Molecular Dynamics. J. Chem. Phys. Scattering from Au(111). J. Phys. Chem. 2019, 151, 070902. Lett. 2017, 8, 666–672. (540) Herr, J. E.; Yao, K.; McIntyre, R.; (531) Huang, S.-D.; Shang, C.; Zhang, X.-J.; Toth, D. W.; Parkhill, J. Metadynam- Liu, Z.-P. Material Discovery by Combin- ics for Training Neural Network Model ing Stochastic Surface Walking Global Chemistries: A Competitive Assessment. Optimization with a Neural Network. J. Chem. Phys. 2018, 148, 241710. Chem. Sci. 2017, 8, 6327–6337. (541) Grimme, S. Exploration of Chemi- (532) Zhou, X.; Nattino, F.; Zhang, Y.; cal compound, conformer, and Reac- Chen, J.; Kroes, G.-J.; Guo, H.; Jiang, B. tion Space with Meta-Dynamics Simula- Dissociative Chemisorption of Methane tions Based on Tight-Binding Quantum on Ni(111) using a Chemically Accu- Chemical Calculations. J. Chem. Theory rate Fifteen Dimensional Potential En- Comput. 2019, 15, 2847–2862. ergy Surface. Phys. Chem. Chem. Phys. (542) Smith, J. S.; Nebgen, B.; Lubbers, N.; 2017, 19, 30540–30550. Isayev, O.; Roitberg, A. E. Less is More: (533) Zhang, Y.; Hu, C.; Jiang, B. Embed- Sampling Chemical Space with Active ded Atom Neural Network Potentials: Learning. J. Chem. Phys. 2018, 148, 241733.

84 (543) Malbon, C. L.; Zhao, B.; Guo, H.; (551) Butler, K. T.; Davies, D. W.; Yarkony, D. R. On the Nonadiabatic Cartwright, H.; Isayev, O.; Walsh, A. Collisional Quenching of OH(A) by H2: Machine Learning for Molecular and A Four Coupled Quasi-Diabatic State Materials Science. Nature 2018, 559, Description. Phys. Chem. Chem. Phys. 547–555. 2020, –. (552) Haghighatlari, M.; Li, J.; Heidar- (544) Xu, X.; Chen, J.; Zhang, D. H. Zadeh, F.; Liu, Y.; Guan, X.; Head- Global Potential Energy Surface for the Gordon, T. Learning to Make Chemi- H+CH4 ↔ H2+CH3 Reaction using Neu- cal Predictions: The Interplay of Fea- ral Networks. Chin. J. Chem. Phys. ture Representation, Data, and Machine 2014, 27, 373–379. Learning Methods. Chem 2020,

(545) Li, J.; Guo, H. Communication: An Ac- (553) Bishop, C. M. Pattern Recognition and curate Full 15 Dimensional Permutation- Machine Learning, 1st ed.; Springer: ally Invariant Potential Energy Surface New York, 2006. for the OH + CH4 → H2O + CH3 Reac- tion. J. Chem. Phys. 2015, 143, 221103. (554) Halama, N. Machine Learning for Tis- sue Diagnostics in Oncology: Brave (546) Jiang, B.; Guo, H. Six-Dimensional New World. Br. J. Cancer 2019, 121, Quantum Dynamics for Dissociative 431âĂŞ433. Chemisorption of H2 and D2 on Ag(111) on a Permutation Invariant Potential En- (555) Bychkov, D.; Linder, N.; Turkki, R.; ergy Surface. Phys. Chem. Chem. Phys. Nordling, S.; Kovanen, P. E.; Ver- 2014, 16, 24704–24715. rill, C.; Walliander, M.; Lundin, M.; Haglund, C.; Lundin, J. Deep Learning (547) Toyoura, K.; Hirano, D.; Seko, A.; Based Tissue Analysis Predicts Outcome Shiga, M.; Kuwabara, A.; Kara- in Colorectal Cancer. Sci. Rep. 2018, 8, suyama, M.; Shitara, K.; Takeuchi, I. 3395. Machine-Learning-Based Selective Sam- pling Procedure for Identifying the (556) Gómez-Meire, S.; Campos, C.; Low-Energy Region in a Potential En- Falqué, E.; Díaz, F.; Fdez-Riverola, F. ergy Surface: A Case Study on Proton Assuring the Authenticity of Northwest Conduction in Oxides. Phys. Rev. B Spain White Wine Varieties using Ma- 2016, 93, 054112. chine Learning Techniques. Food Res. Int. 2014, 60, 230 – 240. (548) Guan, Y.; Yang, S.; Zhang, D. H. Con- struction of Reactive Potential Energy (557) Watanabe, N.; Murata, M.; Ogawa, T.; Surfaces with Gaussian Process Regres- Vavricka, C. J.; Kondo, A.; Ogino, C.; sion: Active Data Selection. Mol. Phys. Araki, M. Exploration and Evaluation of 2018, 116, 823–834. Machine Learning-Based Models for Pre- dicting Enzymatic Reactions. J. Chem. (549) Vargas-Hernández, R. A.; Guan, Y.; Inf. Model. 2020, 60, 1833–1843. Zhang, D. H.; Krems, R. V. Bayesian Optimization for the Inverse Scattering (558) Chen, T.; Guestrin, C. XGBoost: A Scal- Problem in Quantum Reaction Dynam- able Tree Boosting System. Proceedings ics. New J. Phys. 2019, 21, 022001. of the 22nd ACM SIGKDD International (550) Todorović, M.; Gutmann, M. U.; Coran- Conference on Knowledge Discovery and der, J.; Rinke, P. Bayesian Inference of Data Mining. New York, NY, USA, 2016; Atomistic Structure in Functional Mate- p 785âĂŞ794. rials. npj Comput. Mater. 2019, 5 .

85 (559) Ahneman, D. T.; Estrada, J. G.; Lin, S.; (569) Puskorius, G. V.; Feldkamp, L. A. De- Dreher, S. D.; Doyle, A. G. Predict- coupled extended Kalman filter training ing Reaction Performance in C–N Cross- of feedforward layered networks. IJCNN- Coupling using Machine Learning. Sci- 91-Seattle International Joint Conference ence 2018, 360, 186–190. on Neural Networks. 1991; pp 771–777.

(560) Atahan-Evrenk, S.; Atalay, F. B. Pre- (570) Singraber, A.; Morawietz, T.; Behler, J.; diction of Intramolecular Reorganiza- Dellago, C. Parallel Multistream Train- tion Energy Using Machine Learning. J. ing of High-Dimensional Neural Network Phys. Chem. A 2019, 123, 7855–7863. Potentials. J. Chem. Theory Comput. 2019, 15, 3075–3092. (561) Hofmann, T.; SchÃűlkopf, B.; Smola, A. J. Kernel Methods in (571) Behler, J. Atom-Centered Symmetry Machine Learning. Ann. Statist. 2008, Functions for Constructing High- 36, 1171–1220. Dimensional Neural Network Potentials. J. Chem. Phys. 2011, 134, 074106. (562) Raschka, S.; Mirjalili, V. Python Ma- chine Learning, 3rd ed.; Packt Publish- (572) LeCun, Y.; Bengio, Y. The Handbook of ing, 2019. Brain Theory and Neural Networks; The MIT Press, Cambridge, MA, USA, 1995; (563) Xue, B.-X.; Barbatti, M.; Dral, P. O. pp 255–257. Machine Learning for Absorption Cross Sections. ChemRxiv 2020, (573) Krizhevsky, A.; Sutskever, I.; Hin- DOI:10.26434/chemrxiv.12594191.v1 . ton, G. E. ImageNet Classification with Deep Convolutional Neural Networks. (564) Ramakrishnan, R.; von Lilienfeld, O. A. 2012, 1097–1105. Reviews in Computational Chemistry; John Wiley & Sons, Ltd, 2017; Chapter (574) Sainath, T. N.; Kingsbury, B.; Saon, G.; 5, pp 225–256. Soltau, H.; rahman Mohamed, A.; Dahl, G.; Ramabhadran, B. Deep Convo- (565) Bartók, A. P.; Kondor, R.; Csányi, G. On lutional Neural Networks for Large-scale Representing Chemical Environments. Speech Tasks. Neural Networks 2015, Phys. Rev. B 2013, 87, 184115. 64, 39 – 48.

(566) Glorot, X.; Bengio, Y. Understanding (575) Gilmer, J.; Schoenholz, S. S.; Riley, P. F.; the Difficulty of Training Deep Feed- Vinyals, O.; Dahl, G. E. Neural Message forward Neural Networks. Proceedings Passing for Quantum Chemistry. Pro- of the Thirteenth International confer- ceedings of the 34th International Con- ence on Artificial Intelligence and Statis- ference on Machine Learning - Volume tics. Chia Laguna Resort, Sardinia, Italy, 70. 2017; p 1263âĂŞ1272. 2010; pp 249–256. (576) Schütt, K. Learning Representations of (567) Duchi, J.; Hazan, E.; Singer, Y. Adap- Atomistic Systems with Deep Neural tive Subgradient Methods for Online Networks. Doctoral Thesis, Technische Learning and Stochastic Optimization. J. Universität Berlin, Berlin, 2018. Mach. Learn. Res. 2011, 12, 2121–2159. (577) Dral, P. O. MLatom: A Program Pack- (568) Kingma, D. P.; Ba, J. Adam: A age for Quantum Chemical Research As- Method for Stochastic Optimization. sisted by Machine Learning. J. Comput. arXiv 2014, abs/1412.6980, 1412.6980. Chem. 2019, 40, 2339–2347.

86 (578) Christensen, A.; Faber, F.; Machine-Learned Potential Energy Sur- Huang, B.; Bratholm, L.; faces. J. Chem. Theory Comput. 2017, Tkatchenko, A.; Müller, K.; Lilien- 13, 4012–4024. feld, O. QML: A Python Toolkit for Quantum Machine Learning. (586) Unke, O. T.; Meuwly, M. PhysNet: A https://github.com/qmlcode/qml, Neural Network for Predicting Ener- 2017. gies, Forces, Dipole Moments, and Par- tial Charges. J. Chem. Theory Comput. (579) Hansen, K.; Biegler, F.; Ramakr- 2019, 15, 3678–3693. ishnan, R.; Pronobis, W.; von Lilienfeld, O. A.; Müller, K.-R.; (587) Lubbers, N.; Smith, J. S.; Barros, K. Tkatchenko, A. Machine Learning Hierarchical Modeling of Molecular En- Predictions of Molecular Properties: ergies using a Deep Neural Network. J. Accurate Many-Body Potentials and Chem. Phys. 2018, 148, 241715. Nonlocality in Chemical Space. J. Phys. (588) Zheng, F.; Gao, X.; Eisfeld, A. Ex- Chem. Lett. 2015, 6, 2326–2331. citonic Wave Function Reconstruction from Near-Field Spectra Using Machine (580) Pozdnyakov, S. N.; Willatt, M. J.; Learning Techniques. Phys. Rev. Lett. Bartók, A. P.; Ortner, C.; Csányi, G.; Ce- 2019, 123, 163202. riotti, M. On the Completeness of Atomic Structure Representations. arXiv 2020, (589) Fabrizio, A.; Grisafi, A.; Meyer, B.; Ceri- 2001.11696 . otti, M.; Corminboeuf, C. Electron Den- sity Learning of Non-Covalent Systems. (581) Çaylak, O.; von Lilienfeld, O. A.; Chem. Sci. 2019, 10, 9424–9432. Baumeier, B. Wasserstein Metric for Improved QML with Adjacency Ma- (590) Grisafi, A.; Fabrizio, A.; Meyer, B.; trix Representations. arXiv 2020, Wilkins, D. M.; Corminboeuf, C.; Ceri- 2001.11005 . otti, M. Transferable Machine-Learning Model of the Electron Density. ACS (582) Bowman, J. M.; Czakó, G.; Fu, B. High- Cent. Sci. 2019, 5, 57–64. Dimensional Ab Initio Potential Energy Surfaces for Reaction Dynamics Calcula- (591) Fabrizio, A.; Briling, K.; Grisafi, A.; tions. Phys. Chem. Chem. Phys. 2011, Corminboeuf, C. Learning (from) the 13, 8094–8111. Electron Density: Transferability, Con- formational and Chemical Diversity. (583) Herr, J. E.; Koh, K.; Yao, K.; Parkhill, J. CHIMIA Int. J. Chem. 2020, 74, 232– Compressing Physics with an Autoen- 236. coder: Creating an Atomic Species Rep- resentation to Improve Machine Learn- (592) Mai, S.; Plasser, F.; Dorn, J.; Fu- ing Models in the Chemical Sciences. J. manal, M.; Daniel, C.; González, L. Chem. Phys. 2019, 151, 084103. Quantitative Wave Function Analysis for Excited States of Transition Metal Com- (584) Kang, B.; Seok, C.; Lee, J. Prediction plexes. Coord. Chem. Rev. 2018, 361, 74 of Molecular Electronic Transitions us- – 97. ing Random Forests. ChemRxiv 2020, DOI:10.26434/chemrxiv.12482840.v1 . (593) Mennucci, B.; Cappelli, C.; Guido, C. A.; Cammi, R.; Tomasi, J. Structures (585) Richings, G. W.; Habershon, S. Di- and Properties of Electronically Excited rect Quantum Dynamics Using Grid- Chromophores in Solution from the Po- Based Wave Function Propagation and larizable Continuum Model Coupled to the Time-Dependent Density Functional

87 Theory. J. Phys. Chem. A 2009, 113, (603) Kasha, M.; Rawls, H. R.; El- 3009–3020. Bayoumi, M. A. The Exciton Model in Molecular Spectroscopy. Pure and (594) Jasper, A. W.; Kendrick, B. K.; Applied Chemistry 1965, 11, 371 – 392. Mead, C. A.; Truhlar, D. G. Modern Trends in Chemical Reaction Dynamics; (604) Rogers, D.; Hahn, M. Extended- World Scientific, 2004; pp 329–391. Connectivity Fingerprints. J. Chem. Inf. Model. 2010, 50, 742–754. (595) Yarkony, D. R. In conical Intersections; Domcke, W., Yarkony, D. R., Köppel, H., (605) Durant, J. L.; Leland, B. A.; Eds.; Advanced Series in Physical Chem- Henry, D. R.; Nourse, J. G. Reopti- istry; World Scientific, 2004; Vol. 15. mization of MDL Keys for Use in Drug Discovery. J. Chem. Inf. Comput. Sci. (596) Cupellini, L.; Bondanza, M.; Not- 2002, 42, 1273–1280. toli, M.; Mennucci, B. Successes & Chal- lenges in the Atomistic Modeling of (606) Landrum, G. RDKit: Open-Source Light-Harvesting and its Photoregula- Cheminformatics Software. 2016, tion. Biochim. Biophys. Acta, Bioenerg. 2020, 1861, 148049. (607) Jain, A.; Ong, S. P.; Hautier, G.; Chen, W.; Richards, W. D.; Dacek, S.; (597) Richings, G. W.; Robertson, C.; Haber- Cholia, S.; Gunter, D.; Skinner, D.; shon, S. Can We Use on-the-Fly Quan- Ceder, G.; Persson, K. A. Commen- tum Simulations to Connect Molecular tary: The Materials Project: A Mate- Structure and Sunscreen Action? Fara- rials Genome Approach to Accelerating day Discuss. 2019, 216, 476–493. Materials Innovation. APL Mater. 2013, (598) Behler, J.; Lorenz, S.; Reuter, K. 1, 011002. Representing Molecule-Surface Interac- (608) Aarva, A.; Deringer, V. L.; Sainio, S.; tions with Symmetry-Adapted Neural Laurila, T.; Caro, M. A. Understand- Networks. J. Chem. Phys. 2007, 127, ing X-Ray Spectroscopy of Carbonaceous 014705. Materials by Combining Experiments, (599) Behler, J. Dissociation of Oxygen Density Functional Theory, and Machine Molecules on the Al(111) Surface. Ph.D. Learning. Part II: Quantitative Fitting thesis, Technical University Berlin, 2004. of Spectra. Chem. Mat. 2019, 31, 9256– 9267. (600) la Cour Jansen, T.; Rettrup, S.; Sarma, C.; Snijders, J.; Palmieri, P. On (609) Bartók, A. P.; De, S.; Poelking, C.; Bern- the Evaluation of Spin-Orbit Coupling stein, N.; Kermode, J. R.; Csányi, G.; Matrix Elements in a Spin-Adapted Ba- Ceriotti, M. Machine Learning Unifies sis. Int. J. Quantum Chem. 1999, 23–27. the Modeling of Materials and Molecules. Sci. Adv. 2017, 3 . (601) Chen, W.-K.; Fang, W.-H.; Cui, G. A Multi-Layer Energy-Based Fragment (610) Aarva, A.; Deringer, V. L.; Sainio, S.; Method for Excited States and Nona- Laurila, T.; Caro, M. A. Understand- diabatic Dynamics. Phys. Chem. Chem. ing X-ray Spectroscopy of Carbonaceous Phys. 2019, 21, 22695–22699. Materials by Combining Experiments, Density Functional Theory, and Machine (602) Christensen, A. S.; von Lilienfeld, O. A. Learning. Part I: Fingerprint Spectra. Operator Quantum Machine Learning: Chem. Mat. 2019, 31, 9243–9255. Navigating the Chemical Space of Re- sponse Properties. CHIMIA 2019, 73, (611) Deringer, V. L.; Caro, M. A.; Jana, R.; 1028–1031. Aarva, A.; Elliott, S. R.; Laurila, T.;

88 CsÃąnyi, G.; Pastewka, L. Computa- the Protein Backbone. Journal of the tional Surface Chemistry of Tetrahedral American Chemical Society 2010, 132, Amorphous Carbon by Combining Ma- 7769–7775. chine Learning and Density Functional Theory. Chem. Mat. 2018, 30, 7438– (619) Zhang, X.-X.; Würth, C.; Zhao, L.; 7445. Resch-Genger, U.; Ernsting, N. P.; Sa- jadi, M. Femtosecond Broadband Flu- (612) Janet, J. P.; Kulik, H. J. Predicting Elec- orescence Upconversion Spectroscopy: tronic Structure Properties of Transition Improved Setup and Photometric Cor- Metal Complexes with Neural Networks. rection. Rev. Sci. Instrum. 2011, 82, Chem. Sci. 2017, 8, 5137–5152. 063108.

(613) Janet, J. P.; Duan, C.; Yang, T.; (620) Ahmad, I.; Ahmed, S.; Anwar, Z.; Nandy, A.; Kulik, H. J. A Quantita- Sheraz, M. A.; Sikorski, M. Photosta- tive Uncertainty Metric Controls Error bility and Photostabilization of Drugs in Neural Network-Driven Chemical Dis- and Drug Products. Int. J. Photoenergy covery. Chem. Sci. 2019, 10, 7913–7922. 2016, 2016, 1–19.

(614) Janet, J. P.; Gani, T. Z. H.; (621) Mathew, S.; Yella, A.; Gao, P.; Steeves, A. H.; Ioannidis, E. I.; Ku- Humphry-Baker, R.; Curchod, B. F. E.; lik, H. J. Leveraging Cheminformatics Ashari-Astani, N.; Tavernelli, I.; Roth- Strategies for Inorganic Discovery: Ap- lisberger, U.; Nazeeruddin, M. K.; plication to Redox Potential Design. Ind. Grätzel, M. Dye-Sensitized Solar Cells Eng. Chem. Res. 2017, 56, 4898–4910. with 13% Efficiency Achieved Through the Molecular Engineering of Porphyrin (615) Janet, J. P.; Chan, L.; Kulik, H. J. Ac- Sensitizers. Nat. Chem. 2014, 6, 242– celerating Chemical Discovery with Ma- 247. chine Learning: Simulated Evolution of Spin Crossover Complexes with an Ar- (622) Dobson, C. M. Chemical Space and Biol- tificial Neural Network. J. Phys. Chem. ogy. Nature 2004, 432, 824–828. Lett. 2018, 9, 1064–1071.

(616) Janet, J. P.; Ramesh, S.; Duan, C.; Ku- lik, H. J. Accurate Multiobjective De- sign in a Space of Millions of Transition Metal Complexes with Neural-Network- Driven Efficient Global Optimization. ACS Cent. Sci. 2020, 6, 513–524.

(617) Ðorđević, N.; Beckwith, J. S.; Yarema, M.; Yarema, O.; Rosspeint- ner, A.; Yazdani, N.; Leuthold, J.; Vauthey, E.; Wood, V. Machine Learn- ing for Analysis of Time-Resolved Luminescence Data. ACS Photonics 2018, 5, 4888–4895.

(618) Abramavicius, D.; Jiang, J.; Bul- heller, B. M.; Hirst, J. D.; Mukamel, S. Simulation Study of Chiral Two- Dimensional Ultraviolet Spectroscopy of

89