Interfacial Potentials in Ion Solvation
A dissertation submitted to the
Graduate School of the University of Cincinnati
in partial fulfillment of the requirements
for the degree of
Doctor of Philosophy
in the Department of Physics
of the McMicken College of Arts and Sciences
by
Carrie Conor Doyle
B.S. in Physics, Rutgers, the State University of New Jersey, 2013
June 2020
supervised by
Dr. Thomas L. Beck
Committee Co-Chair: Dr. Carlos Bolech, Physics
Committee Member: Dr. Rohana Wiedjewardhana, Physics
Committee Member: Dr. Leigh Smith, Physics Abstract
Solvation science is an integral part of many fields across physics, chemistry, and biology.
Liquids, interfaces, and the ions that populate them are responsible for many poorly understood natural phenomena such as ion specific effects. Establishing a single-ion solvation free energy thermodynamic scale is a necessary component to unraveling ion-specific effects.
This task is made difficult by the experimental immeasurability of quantities such as the interfacial potential between two media, which sets the scale. Computer simulations provide a necessary bridge between experimental and theoretical results. However, computer models are limited by the accuracy-efficiency dilemma, and results are misinterpreted when the underlying physics is overlooked. Classical molecular dynamic techniques, while efficient, lack transferability. Quantum-based ab initio techniques are accurate and transferable, but their inefficiency limits the accessible simulation size and time. This thesis seeks to determine the physical origin of the interfacial potential at the liquid-vapor interface using classical models.
Additionally, I assess the ability of Neural Network Potential (NNP) simulation methods to produce electrostatic properties of bulk liquids and interfaces. Complicating factors are minimized through a simple water model (SPC/E) free of experimental parametrization and a
finite droplet simulation free of Ewald effects. Multipolar decomposition of the potential in the region of zero charge density provides a direct method for determining the potential felt by ions near interfaces. Non-aqueous solvents are studied through an OPLS-AA based model of the organic liquid ethylene carbonate (EC) using the same approach to compare aqueous and non-aqueous solvents. Neural network potentials may be a step towards the higher-level
1 need for predictive models, but they require further testing. Using an existing NNP framework, I train models for dynamics, as well as multipole electrostatics. Combining the two in both a bulk and interfacial system allows for the calculation of interfacial electrostatic properties. My results for water elucidate the reason for widely varying net potentials calculated for various models with similar dipole but differing quadrupole moments.
Near-cancelling dipole contributions between the droplet interface and the cavity interface of a solvated ion leaves the quadrupole as the dominant contribution to the net potential.
Molecular density profiles and potential profiles show that a length scale of 5 Å from cavity boundary is needed for a convergent potential. A theoretical argument for the radial dependence of each contribution is made, which supports my results. EC also has this radial dependence but has a different length scale of convergence. Differences in the molecular size, orientation, hydrogen-bonding capabilities, and multipole moments results in solvent-specific net potential contributions. This is evidenced by the results for charged cavities and an orientational analysis of EC. NNPs are shown to provide excellent agreement with ab initio electrostatic properties. This is encouraging evidence for the use of NNPs in the calculation of thermodynamic properties and in force field development.
2
Acknowledgements
I would like to express my gratitude to all those who were a part of my PhD studies. The
guidance and support of my advisor Dr. Thomas L. Beck was an essential part of my thesis work. His vast interdisciplinary knowledge and encouragement of independent research was a
constant source of motivation. I thank Dr. Bolech, Dr. Wiedjewardhana, and Dr. Smith for
their contributions as my committee members, and as professors, along with others at the
University of Cincinnati and Rutgers University who prepared me for rigorous scientific
investigations.
Lab group members Zohre Gorunmez, Mimi Liu, and Andrew Eisenhardt were always there
to commiserate and grow with. Extensive discussions and collaboration with Yu Shi were an
important part of the evolution of my research and a fond memory. I am forever grateful for
the friends I met along the way.
Finally, I thank my family, Lorraine, William, and Jessie Doyle for the constant support and
encouragement, no matter the distance between us.
4 Yes, as everyone knows, meditation and water are wedded for ever. —Herman Melville
Dedicated to William Doyle
5 Contents
1 Introduction 13
1.1 Preface...... 13
1.2 Evolution of the Computer Simulation of Liquids...... 16
1.2.1 History...... 17
1.2.2 Connection with Theory and Experiment...... 18
1.3 Solvation Science...... 20
1.3.1 Mixtures and Phase Separation...... 20
1.3.2 Proteins and Membranes...... 22
1.3.3 Specific Ion Solvation: Hofmeister Series...... 24
1.4 Theories and Simulation of Ion Solvation...... 27
1.4.1 Past...... 27
1.4.2 Present...... 30
Classical Simulation...... 30
Ab Initio Simulation...... 32
Neural Network Potentials...... 33
1.5 Interfacial Potential Effects...... 35
1.5.1 Thermodynamic Scale...... 35
1.5.2 Are they measurable?...... 38
6 CONTENTS
1.5.3 What are we measuring?...... 41
1.6 Summary...... 45
2 Theory 48
2.1 Preface...... 48
2.2 Classical Molecular Dynamics...... 48
2.2.1 Equations of Motion...... 49
2.2.2 Thermodynamic Ensembles and Equilibration...... 50
2.2.3 Classical Force Fields...... 50
2.2.4 SPC/E Force Field...... 52
2.2.5 OPLS-AA...... 53
2.2.6 Boundary Conditions...... 53
2.3 Ab Initio Molecular Dynamics...... 54
2.4 Neural Network Potentials...... 56
2.5 Thermodynamics of Ion Solvation...... 59
2.5.1 The Potential Distribution Theorem...... 60
2.5.2 Quasichemical Theory...... 61
2.5.3 Interfacial Potentials...... 63
2.6 Macroscopic Interfacial Electrostatics...... 65
2.6.1 Multipole Expansion of a Charge Distribution...... 66
2.6.2 Coordinate Systems: Cartesian and Spherical...... 67
2.6.3 Molecular Multipole Expansion of Electrostatic Potential...... 70
3 Water Liquid-Vapor Interfacial Potential Shifts 71
3.1 Preface...... 71
7 CONTENTS
3.2 Computational Methods...... 73
3.3 Results and Discussion...... 74
3.4 Conclusion...... 84
4 Ethylene Carbonate Liquid-Vapor Interfacial Potential Shifts 87
4.1 Preface...... 87
4.2 Computational Methods...... 90
4.3 Results and Discussion...... 91
4.3.1 Multipole moments of EC and Water...... 91
4.3.2 Electrostatic Potential Analysis: Neutral Cavity...... 92
4.3.3 Electrostatic Potential and Molecular Orientation...... 97
4.3.4 Electrostatic Potential Analysis: Charged Cavities...... 103
4.4 Conclusion...... 108
5 Electrostatic Properties from Neural Network Potentials 112
5.1 Preface...... 112
5.2 Computational Methods...... 114
5.2.1 Ab Initio Simulation...... 114
5.2.2 DeepMD reconstruction of the PES...... 117
5.2.3 DNN for determination of Multipole Moments: DeePD, DeePM, DeePPM 119
5.2.4 DNN for determination of DeePMD multipole moments...... 120
5.3 Results and Discussion...... 120
5.4 Conclusion...... 129
6 Conclusions 131
8 CONTENTS
A Appendix 153
A.1 Molecular Multipole Definition...... 153
A.2 Supporting Information: Chapter 3...... 155
A.3 Supporting Information: Chapter 4...... 159
9 List of Figures
1.1 Diagram of computational simulations, theory, and experiment...... 19
1.2 Schematic of boundary conditions and interfacial potential shifts...... 45
2.1 Schematic of basic neural network...... 57
2.2 Schematic of studied system- a liquid solvent droplet and interfacial potential
shifts...... 64
3.1 SPC/E electrostatic potential profile for different cavity sizes...... 76
3.2 SPC/E electrostatic poential multipole expansion and RDF for differnet cavity
sizes...... 78
3.3 SPC/E net potential dependence on cavity size: multipole decomposition.... 83
4.1 EC electrostatic potential profile for different cavity sizes...... 93
4.2 EC electrostatic potential multipole expansion and RDF for different cavity sizes 95
4.3 EC net potential dependence on cavity size: multipole decomposition...... 96
4.4 Diagram of EC orientation...... 99
4.5 EC distribution of dipole vector projection, neutral cavity...... 101
4.6 EC distribution of planar angle, neutral cavity...... 103
4.7 EC electrostatic potential profile for different cavity sizes, charged cavities... 106
4.8 EC distribution of planar angle, positive charge cavity...... 107
10 LIST OF FIGURES
4.9 EC distribution of planar angle, negative charge cavity...... 108
5.1 Schematic of DNN framework for electrostatic properties...... 115
5.2 Loss functions for DeePMD, DeePD, DeePM, DeePPM...... 121
5.3 Radial Distribution Function of AIMD and DeepMD bulk water...... 122
5.4 Dipole moment distribution for AIMD and DeePMD bulk water...... 123
5.5 Eigenvalues of the primitive quadrupole moment for AIMD and DeePMD bulk
water...... 124
5.6 Trace of the primitive quadrupole moment for AIMD and DeePMD bulk water. 125
5.7 Distribution of the double dot product of the primitive quadrupole moment for
AIMD and DeePMD bulk water...... 126
5.8 Radial Distribution Function of AIMD and DeepMD interfacial cavity system.. 127
5.9 Average dipole moment of for AIMD and DeePMD Interfacial Cavity System in
shells around cavity...... 128
5.10 Average traceless quadrupole double dot product for AIMD and DeePMD Inter-
facial Cavity System in shells around cavity...... 129
11 List of Tables
3.1 Electrostatic Contributions to Net Potential at Different Cavity Radii. Poten-
tials Computed at the Cavity Center with Contributions from Cavity-Liquid and
Liquid-Vapor Interfaces a ...... 79
3.2 Electrostatic Potential Shifts from the Cavity-Liquid and Liquid-Vapor Interfaces
at Different Cavity Radii. a ...... 80
4.1 Quadrupole Components of ethylene carbonate (EC) and water (SPC/E) a ... 92
4.2 Electrostatic Contributions to Net Potential at Different Cavity Radii. Poten-
tials Computed at the Cavity Center with Contributions from Cavity-Liquid and
Liquid-Vapor Interfaces a ...... 98
4.3 Electrostatic Contributions to Net Potential for Charged Cavities. a ...... 104
12 Chapter 1
Introduction
1.1 Preface
Life would not exist without the fascinating compound known as water. Society would not have developed without the ability to harness energy from the natural world for machines. Water and energy are unequivocally important for understanding the physical phenomena on earth, in our bodies, and to power civilizations, and yet there is still a lot to be learned about them.
Water, which composes 99% of the molecules in the human body,1 and 71% of the Earth’s surface2 is still a mystery. Energy, and improving ways to generate and store it, is a scientific problem which has grown exponentially as demands for energy increase.
It is hard to separate the study of water or the improvement of energy storage devices (bat- teries) from salts. These ion pairs exist in varying concentrations in almost all of the water on earth from seawater to the cytoplasm in your cells. They drive biological functions and help to regulate earth’s atmosphere. Since the advent of galvanic cells, energy has been stored in batteries through ions. Quantitative and predictive models of ion solvation are critically impor- tant to the improvement of energy storage devices and understanding of physical phenomena
13 1.1. PREFACE in a wide range of scientific research.
The study of liquid solvation was forever changed by the incorporation of computer sim- ulations in the mid-20th century. By introducing a new category of science— simulations— theories could be tested and experiments validated, while giving access to physical parameters which could not be achieved through experiment. Now, with the newest revolution, that of machine learning, computers are again poised to change the way that we study liquid systems.
With physically grounded principles these new methods can be used to access length scales and time scales which are drastically limited in quantum ab initio simluation methods. With these tools, I tackle the problem of ion solvation with the goal of improved physical understanding of molecular liquids. While accuracy in measurement is important for some purposes, I aim to answer the fundamental question of the origin of the interfacial electrostatic potential and whether or not neural network potentials can be used to study the electrostatics of ion solvation.
In the third chapter of this thesis I use the simplest toy model for a water molecule. Stripping away the decades of refinement in various water models which are fit to experimental results, I use classical MD to investigate the multipolar origin of the interfacial potential at the surface of liquid water. I implement a new way of analyzing the components that is more intuitive than previous methods, and leads to a quick realization that the quadrupolar potential close to the solvated ion is the dominant contributor to the potential which the ion feels. This result solves the long-standing question of why various types of water models produce varying interfacial potential values. Additionally, this analysis established a length scale at which electrostatic potentials are physically meaningful— where the charge density averages to zero.
I then extend this analysis to ethylene carbonate, an organic liquid relevant to energy stor- age, and which has a range of academic interest behind it. A popular solvent in lithium-ion batteries, this carbonate and the related propylene carbonate and dimethyl carbonate are in-
14 1.1. PREFACE teresting because with just a small structural change in the atomic constituents, many different chemical properties are changed. Using similar methods to those I used for water, I investi- gate whether additional insights can be gained. Are there any general properties of interfacial potentials that exist in these differing solvent molecules? Indeed, I find that there are length scales at which the interfacial potential is physically meaningful in ethylene carbonate, though the specific multipolar contributions to the potential are quite different. Using an orientational analysis, I find that the sign of the quadrupolar potential is dependent on the quadrupole mo- ment, the orientation of the molecule, and surface geometry. While general principles exist, knowledge of charge symmetries contained within a molecule’s unique multipole moments and the orientation of the molecule at the interface produce the interfacial potential.
Throughout this work, I highlight the limitations of classical and quantum simulation meth- ods. While I employ classical MD in my first two projects, it is used to extract understanding of general physical principles, not to calculate highly accurate results. In my final project, I used the emerging field of Neural Network Potentials to simulate liquid systems of ab initio accuracy at classical cost. Validating these methods for use in the study of the electrostatic properties of liquids is a necessary step in using these models for the calculation of physical properties.
The meaningful impact of this thesis is to establish and clarify the physical origin of the interfacial potential in solvents. While theoretical and experimental methods have started to zero in on a value of -0.4V for liquid water, the questions remain of how this value is generated, and what will it be for other solvents. The importance of system size is demonstrated, as it establishes a zeroing of the potential, a concept that is largely ignored in many accepted works.
New methods of Neural Network Potential are shown to be successful and exciting new methods in the science of liquid simulation.
15 1.2. EVOLUTION OF THE COMPUTER SIMULATION OF LIQUIDS
1.2 Evolution of the Computer Simulation of Liquids
Physics is the study of the universe through substances and their interactions. We learn about these interactions though the equations that describe them. Prior to the 20th century, knowl- edge of the physics of materials and many-body systems was restricted to two categories.
One category sought to solve these equations analytically and exactly. In this case, there are huge limits on what systems can be modeled. Newton’s Equations of Motion, for example, are restricted to solving for two interacting objects. Materials are composed of many more than two objects. Avogadro’s number tells us that one mole of a substance contains 6.022 x 1023 molecules. For these many-body systems there are, again, a very limited amount of problems for which equilibrium properties can be solved for exactly. Namely, the Ising Model of Ferromagnetism, the Ideal gas, and a harmonic crystal.3 The Ideal Gas Law can be derived directly from the Kinetic Theory of Gases and the Maxwell-Boltzmann Distribution using statistical mechanics.
The other category uses approximate equations that describe the macroscopic, large-scale behavior of a substance. These theories ignore the intricacies of the underlying microscopic behavior of the constituent atoms and molecules and their interactions, in favor of a more qualitative description of how a substance behaves. Limited knowledge of the intermolecular interaction leads to equations which attempt to describe the physics. Examples include the
Van der Waals Equation for dense gases, Debye-Hückel theory, and the Boltzmann Equation for intermolecular interactions.3 While these theories expand the number of systems we can study, they are still approximations. If theoretical and experimental results disagree, we are left with the question of whether the theory itself is wrong, or our assumptions about the intermolecular interactions are wrong. This question is where computer simulations had a fortuitous entrance in the scientific timeline.
16 1.2. EVOLUTION OF THE COMPUTER SIMULATION OF LIQUIDS
1.2.1 History
The simulation of liquids using computers arose out of scientific necessity and chance. Prior to 1905, there was no physical explanation for the seemingly random motion of dead pollen molecules suspended in a liquid, termed Brownian motion after the botanist who studied the phenomena. Unexplained by liquid currents, evaporation, or Maxwell-Boltzmann Kintetic
Theory of Gases, Einstein and Smoluchowski independently reached similar conclusions about
Brownian Motion. That is, large suspended molecules (pollen) experience motion due to their collisions with the small molecules of the liquid (solvent)— and further that mean displacement and observation time are proportional to the Diffusion Constant.4 This novel result not only settled the idea that liquids have a non-continuous behavior, but also marked the first time that an experimental observable in liquid could be connected with theory. Later works by Morrell and Hildebrand (1936) compared liquid mercury x-ray diffraction data to their 3D gelatin ball model, while examining factors such as temperature on intermolecular distance.5
There are obvious limitations in these types of experiments. Suspending particles in liquids or building physical ball models while simple and intuitive, are not suitable to handle the large scale collective properties that must be calculated for liquids. The integration of programmable digital computers into military and rare scientific operations by the 1930s and 1940s was an auspicious occurrence which has altered the trajectory of liquid science. Computers, which have the capacity to store large quantities of information that well exceeds that of the human brain, perform operations on data, and incorporate randomness, are perfect candidates to handle liquid experiments. This led to the first computer simulation of a liquid state in 1953 by Metropolis et al.6 at Los Alamos National Laboratories7 on the MANIAC (Mathematical
Analyzer Numerical Integrator and Computer)— known as "Monte Carlo" simulation.
The original Monte Carlo algorithm works to choose configurations from a probability distri-
17 1.2. EVOLUTION OF THE COMPUTER SIMULATION OF LIQUIDS bution. After this landmark paper, the field developed to include Molecular Dynamics (MD)— which aims to solve Newton’s classical equations of motion for a system of particles. Throughout the 1970s and 1980s, researchers incorporated increasingly complicated molecules and methods into MD. It is important to remember for what reason these methods were being developed and how simulations are situated in the hierarchy of science.
1.2.2 Connection with Theory and Experiment
The underpinning of Molecular Dynamics and other liquid simulation methods lies in the idea that many desired physical quantities can be calculated if the dynamics of many interacting particles can be determined. The entire system’s coordinates ~x and momenta ~p (or velocities), once solved for, allow us to determine energies, entropies, and other thermodynamic or transport properties. Statistical Mechanical Theory allows us to use an array of trajectories to calculate averaged quantities. However, just as a computer program is only as correct as the computer programmer, these results are only as physically correct as the models used in simulation.
The usefulness of simulations is their ability to connect experimental results with theoretical predictions as shown in Figure 1.1. Simulation results can be compared to results of approximate theories, providing tests for theories, or to experimental results, providing evidence for or against the model used. If the model does work well, it can be used to access timescales (both large and small), temperatures, pressures, and any other parameter that is impossible to measure with an experiment.
18 1.2. EVOLUTION OF THE COMPUTER SIMULATION OF LIQUIDS
Liquids Model Liquids
Do Computer Make Approx. Simulations Theories
Experimental Exact Results Theoretical Results for Model Predictions
Compare Compare
Tests of Tests of Models Theories Figure 1.1: Diagram relating computational simulation to theory and experiment in the science of liquids. Figure inspired by Ref.7, diagram otherwise original.
In this thesis I use simulation largely as a tool for theoretical prediction. As I will explain below, electrostatic properties of liquids are difficult to access with both experiment and theory.
Simulation models are extremely useful in electrostatics for a few reasons. First, the additive nature of charge, fields, potentials, etc. is perfectly suited for simple ensemble averaging.
Additionally, the basic theory of electrostatics has been understood for centuries. Finally, it is possible to extrapolate simple toy-models of interacting charges to slightly more complicated or more realistic molecular structures—allowing for predictions beyond the original simulation model.
19 1.3. SOLVATION SCIENCE
1.3 Solvation Science
The science of liquids and simulations have been intertwined since the advent of computers.
One field is Solvation Science, or the study of ionic and molecular liquids and interfaces. En- compassing much of physical chemistry, biochemistry, atmospheric science, and other diverse areas, the study of Solvation Science has wide ranging applications. Here, I will discuss a few of the major scientific questions. After presenting the broad goals of the fields, I will discuss how current simulation models are proving unequipped to solve problems outside of the areas for which they were originally developed. Connecting to Figure 1.1, the comparison of model results to experiment is not favorable. In section 1.4, I will discuss why these models fail.
First, I will discuss fluid mixtures and phase separation. Then I will summarize some relevant research in membrane proteins. Finally, I will introduce the study of specific ion solvation. While these are only a select few areas of research, they are all difficult to model with existing theories, thus providing adequate motivation for this thesis.
1.3.1 Mixtures and Phase Separation
Classical simulation models for water are generally parametrized for unique states of a liquid at ambient temperature. Complicated many-body effects are squeezed into a few parameters within a given model. The experimental gas phase dipole moment of a water molecule is 1.85 D, but the accepted values for water and ice range from 2.5 to 3.0 D. The simplest computational water models do not incorporate this fluctuation between phases, instead they use average values— the SPC/E classical model of water has a dipole of 2.36D. Alternative polarizable models such as dipole-polarizable or fluctuating-charge models incorporate polarizability as a response to the local electric field. These models offer improvement for a slightly higher computational cost. Polarizabilty is just one element needed for an accurate model of different
20 1.3. SOLVATION SCIENCE phases of a solution.
One application affected by these shortcomings is phase diagrams. Phase diagrams show the thermodynamic parameters or concentrations at which different phases of matter exist. At phase boundaries, the two phases coexist. The classic example is the Pressure v. Temperature plot for solid, liquid, and gaseous water. For systems of more than one substance, the Liquid-
Liquid Equilibrium defines the concentrations and conditions at which two or more liquids are at equilibrium. Knowledge of these concentrations are beneficial to advance the understanding of pharmokinetic properties of drugs in different living biophases, or pollutant molecule behavior in organic environments such as soil and groundwater.89 Environmental chemistry, drug design, and energy storage are some areas of research where phase separation is vitally important.
Simulations are a large component of phase separation research, because determining phase boundaries for particular mixtures at certain conditions requires extensive experiments. Ap- proaches such as SAFT10 (Statisical Associating Fluid Theory) rely on experimental data. The difficulty is being able to predict these boundaries for previously undetermined combinations of liquid compounds at specific conditions, without experiment. Some studies have proven the ability of certain force field models to accurately produce phase behavior for some liquids— wa- ter and alcohol (n-butanol) for example.11 Despite this, the challenges of simulating adequate conformational space and obtaining properly parametrized force fields for the liquids remains.
One study of LLE for Dipropylene glycol dimethyl ether and water pointed to the possibility of the system not reaching equilibria, inadequate sampling of compositional fluctuations, and small system size for deviation of mutual solubilities with experiment.12
One attempt to improve the modeling of water in phase separation added coupling of
Lennard-Jones repulsion parameters to the magnitude of the fluctuating charge in order to model the charge dependence of the overlap integrals in quantum mechanics. Although certain
21 1.3. SOLVATION SCIENCE properties such as vapor and liquid densities, heats of vaporization, and dielectric constants showed better agreement with experiment, the models still failed to produce the vapor-liquid coexistence curve at room temperature.13
Clearly there is a need for improved liquid simulation models in order to find predictive mod- els for Liquid-Liquid Equilibria. Mixtures of organic carbonates such as Ethylene Carbonate
(EC) and Dimethyl Carbonate (DMC) are commonly used to solvate LiP F6 in Lithium-ion bat- teries.14 Energy technology benefits from detailed knowledge of the chemistry of these mixtures in order to optimize important metrics such as electrolyte viscosity, melting point, and conduc- tivity. Simulations can provide insights and resolve discrepancies between experiments such as the number of EC or DMC molecules in the solvation shell around the lithium-ion, which differs between electrospray ionization mass spectroscopy and NMR spectroscopy.14 Models which are transferable and incorporate higher-level theory such as polarization and many-body effects are expected to produce accurate phase boundaries.
1.3.2 Proteins and Membranes
All cells and the organelles within them are composed of liquids, the cytoplasm, surrounded by a membrane. This membrane is made of a lipid bi-layer embedded with special proteins which allow the inside of the cell to communicate with it’s environment.15 Signal transduction through ions, catalysis of enzyme reactions, and transport of small molecular nutrients are among some functions of these proteins. These processes are essential to nearly all physiolog- ical activities such as muscle contractions through nerve and muscle excitation, brain activity
(learning and memory), immune response, blood pressure regulation, and hormone secretion.15
Defects and mutations of ion channel proteins are responsible for many different diseases. They provide insights into the functions of ion channels. For example, Cystic Fibrosis is caused
22 1.3. SOLVATION SCIENCE by a reduction of the epithelial Cl- transport. Familial hemiplegic migraines, LQT syndrome
(cardiac disorder leading to sudden death in young people), and Diabetes Mellitus are just a few types of clinical diagnoses that result from genetic mutations generating faulty membrane proteins. Understanding of the physics of these proteins can help researchers develop targeted therapies.16
Membrane proteins are loosely defined by three categories: ion channels, pumps, and trans- porters. Channels are gated pores whose opening and closing is governed by "chemical’ means, voltage gradients, or "mechanical means" through ligand binding. The gating of these channels leads to the ionic composition outside of cells differing from that within cells. Transporters change diffusion of an ion or substrate through existing gradients across the membrane.17 This creates an electrical signal that travels rapidly in excitable cells. The characterization of trans- port mechanisms is an essential step towards understanding their function and in developing therapeutic substances. Thermodynamics governs the transport of ions and the binding prop- erties of ligands which allow for transport.
Simulations are commonly used to understand the thermodynamics of this process after the protein’s crystal structure is determined. In one study of the Gramicidin channel, which could be used in fuel cell technology, the free energy barrier of K+ agreed with experimental results, but the entropy and enthalpy failed. This failure can be attributed to problems with the simulation model used.18 A later study looked at the free energy profiles of K+ and Na+ through the channel using two different models— one including polarization of the solvating molecules using the AMOEBA model, and one without. Including polarization and multipole terms on the atoms, brought the simulation results much closer to experimental free energy results.19
One final example is the CLC chloride channel/transporter, specifically CLC-ec1 of E. coli.
23 1.3. SOLVATION SCIENCE
This channel exchanges a proton for two chloride ions. In order to better understand the mechanics of ion transport Chen et. al.20 used simulations to calculate the absolute hydration free energy of ions in water and on the protein binding sites. It was found through analysis of free energies for gate opening and closing that proton binding is necessary before Cl– can bind to the central binding site, in agreement with experiments. One outstanding question with this channel is the perplexing experimental result that the substitution of F- for Cl- in this transporter effectively stops the flow of protons.21 Classical simulation results do not explain this ion selectivity, although quantum chemistry calculations suggest that the higher proton affinity and smaller size of F- make the passage of the proton more difficult.20 This unresolved issue illustrates the need for more advanced, predictive models for liquid simulation. Additionally, this is a classic example of ion specificity. A "small" change in ion identity generates a substantial change in protein function. The next section will highlight a few specific ion solvation problems.
1.3.3 Specific Ion Solvation: Hofmeister Series
A brief summary of the field of specific ion effects is difficult because of it’s breadth. Historically, the field developed 150 years ago after Franz Hofmeister, a pharmacologist, and his laboratory in
Prague, Czechoslovakia, discovered and published many papers on the ordered ability of certain salts to affect many different physical chemistry phenomena such as dissolving, precipitating, and lycotropic swelling of proteins. The most well-known example is the ordering of anions species (of like charge) by increasing ability to precipitate egg-white protein from solution.
At that time, the scientific fields of pharmacology, botany, biology, and physical chemistry were closely related.22 Scientifically, while these fields have since become segregated, these effects apply to just as many diverse fields as in Hofmeister’s time. Fundamental and applied chemistry, biology/ biophysics, biochemistry, materials science, atmospheric/marine science,
24 1.3. SOLVATION SCIENCE energy, climate science, physiology, colloid/surfactant science, and many other fields could all benefit from improved understanding of specific ion effects and solvation. If the field relates to aqueous (or nonaqueous) liquids, there are related specific ion effects, which in light of practicality and the lack of theoretical understanding, were largely ignored until recently.23
Specific ion effects are the varying abilities of certain ionic species to generate physical phenomena. Extensive theoretical work has not completed the puzzle. Beyond empirically derived rules, there has not been a quantitatively predictive theory developed for the phenom- ena. Questions such as whether ion size, charge density, charge distribution, geometry, and interaction beyond simple electrostatics such as polarizability play a role, all still persist. The goal is to extract generalities and understand the interactions that generate the effects so that we can comprehend the processes of the living systems and develop technologies.23 Examples of aqueous ion specific effects include:
• the alteration of ionic surfactant phase behavior with counter ion identity changes24 25
• accumulation of ions at the water liquid-vapor interface26 27 28
• activity coefficients23 29
• pH measurements29
• protein denaturation30
• phospholipid aggregation23
• surface tension31 32 33
Non aqueous examples, relevant to this thesis in organic carbonates include:
• salts effecting structuredness in liquids34
25 1.3. SOLVATION SCIENCE
• solubility35
• Hofmeister series ordering (not consistent with that in water)36
• bubble-bubble coalescence
Atmospheric science is one area where specific ion effects, in particular those which contain interfacial effects, are wide-spread. The study of the physics and chemistry of clouds, gases, and aerosols is important for understanding the climate of earth and other planets, cloud formation, and the weather.26 The natural world is full of aqueous interfaces which contain ions (like seawa- ter) and the delicate interactions thereof are responsible for many natural phenomena. Liquid and solid aerosols such as fog or clouds and ice clouds or particulates, respectively, are colloids.
The potential of mean force between two colloidal particles in a salt solution has been shown to contain ion-specific dependence37 in addition to charge and size dependence through dispersion forces. Similar ion-specificity has been found in the stability and formation/deformation of amphiphiles such as micelles and biological membranes.38
Theoretical formulations of ion specific effects use ions embedded in a dielectric continuum with ion charge, size, and dielectric constant as variables in the free energy of hydration. In
1.2 I discussed the limitations of macroscopic theories in describing what is clearly microscopic.
Here, the granular nature, local electrostatic effects, and quantum forces near an ion necessitate the use of model liquids beyond approximate models.26 In the next sections I will discuss both past and present limitations to theories of ion solvation, as well as the importance of interfacial potential effects in single-ion solvation free energy calculations necessary for unraveling specific- ion effects.
26 1.4. THEORIES AND SIMULATION OF ION SOLVATION
1.4 Theories and Simulation of Ion Solvation
Early theories of ion solvation developed to explain a deviation from some idealized problem which is formed to make the problem tractable. They ascribe qualities to the system which allow for approximate physical models, such as the solvent as a simple dielectric continuum and the ions as charged point particles or hard spheres. The addition of computer simulations into the picture in the early 1960s, as I discussed in section 1.2, allowed for these theories to be tested using MD and MC. The advantages of simulations is that they do not use macroscopic parameters that overlook the microscopic properties of the system— orientations, ion distribu- tions, hydrogen bonding, etc.— but derive properties from atomistic interactions. These "past" theories are still in use and can be accurate under certain conditions. There are three main categories of simulation in the present day: those using classical MD, ab initio theories, and
Machine Learning derived potentials. Various combinations also exist, but I will focus on them separately, providing a quick introduction to the benefits and drawbacks of each.
1.4.1 Past
Electrolytes are any ionic solid that can dissolve in polar liquids (like water), dissociate into ions and carry an electric current.39 The textbook definition of the electrochemical potential for an ideal electrolyte (neutral salt) in solution with activity coefficient of 1 is,
µ(x) = µ0 + kT ln c(x) + zeψ(x)
where µ0 is the chemical potetial, x is position and ψ(x) the potential at that location. This description is incomplete if the ions are near large charged surfaces such as colloids or have unequal concentrations. Possion-Boltzmann and Debye-Hückel Theory were developed as a
27 1.4. THEORIES AND SIMULATION OF ION SOLVATION theoretical explanation for this deviation, using electrostatic forces as the origin. These theories use the Boltzmann relation— where local average electric potentials define the difference in potential energy between ions— and a Boltzmann distribution in the Poisson Equation,
ρ ∇2φ = e (1.1) r0
where ρe is a local charge density, r is the solvent dielectric constant. Ion motion is governed by Boltzmann statistics and ion density is given by,
−Wi 0 k T ci = ci e B (1.2)
where the Wi = qφ is the work to move an ion towards the surface, ci is the local ion concen-
0 tration, and ci is the bulk ion concentration.
Debye-Hückel theory gives the activities of fully dissociated electrolyte solutions. It relies on several assumptions: The ions are spherical with no polarization, the solvent medium is a dielectric continuum, and the solute is completely dissociated.39 The derivation involves using the Boltzmann distribution for the charge density in Poisson’s Equation. The extra
(electrostatic) free energy per ion is then
κq2 µ = − i (1.3) es 2
−1 2 4π P 2 where κ is the inverse Debye wavelength and κ = kT i qi ci. Essentially, the ions are treated as non-interacting, screened point charges. Basic physics such as dispersion, solvent interaction and polarization are completely ignored. Setting this equal to kT ln(γi) where γi is the activity coefficient gives a measurable way to compare to experiment.
28 1.4. THEORIES AND SIMULATION OF ION SOLVATION
Even so, this theory is only valid at low concentration (100mM/L),40 whereas many specific ion effects emerge beyond this limit. In addition, many systems in the natural world exceed
+ + 2+ 2+ 2+ − 2− − this. Ions such as Na , K , Mg , Ca , CO3 , HCO3 , SO4 , Cl are responsible for maintain- ing the processes of life. In seawater, Na+ and Cl− concentrations often exceed 500 mM/L. In
Mammalian cells (intracellular and extracellular), blood plasma, and interstitial fluid composi- tions are often higher than 100 mM/L.41 Furthermore, recent experiments of aqueous NaCl and other liquids, have shown the opposite of what Debye-Hückel predicts— showing an increase in screening length with increasing concentration.42
Derjaguin-Landau-Verwey-Overbeek (DLVO) theory describes the interaction between charged surfaces (colloids) in a liquid through a screened Coulomb (Yukawa) potential. The interaction is a balance between repulsive entropic and attrative Van der Waals dispersion. Like Debye-
Hückel theory it assumes ions are charged hard spheres in a dielectric continuum. Because polarization of the electron cloud and ion size are ignored, ion specificity is explicitly ignored.43
Other assumptions, such as an ideally smooth surface have been revised to include surface roughness,44 but the general assumptions remain. This model was revised by Onsager and
Samaras33 in 1934 to predict surface tension.
Onsager-Samaras Theory predicts that ions will be expelled from the interface,26 demon- strating how continuum models fail in some theoretical predictions. Beginning with the Gibbs-
Duhem equation for a flat surface and and dividing by surface area A while holding T constant leads to the Gibbs adsorption isotherm where ni is the excess number of moles for species i,
ni Γi = A and surface tension is γ:
dγ = −Γ1dµ1 − Γ2dµ2 (1.4)
For inorganic salts (alkalide haldide series), as measured by Heydweiller31 in the 1910s, there
29 1.4. THEORIES AND SIMULATION OF ION SOLVATION is an increase in surface tension with increased concentration which leads to a negative Gibbs surface excess. This means there should be a depletion of these ions at the interface. Theories of
Wager32 and others33 conceptualized this depletion as a simple fictitious image charge force.26
In reality, this depletion is not seen by simulations or experiments of large polarizable ions, which calls in to question the validity of the entire theory.
The Born model45 of ion solvation is another continuum solvation method used to evaluate the free energy change from vapor (vacuum) to bulk (can be generalize to other mediums). For
46 ion charge q, radius r0, and the dielectric constant of the medium r, the Born Equation is,
N q2 1 µ = − A 1 − (1.5) 8π0r0 r
This approximation is moderately successful, especially for distant ion-solvent interactions. It does account for ion size, although it is assumed they are perfectly spherical, and can be used to judge general behavior of ions in solvents. Still, it ignores the free energy changes of nearby solvent, solvent reorganization, and polarisation and is unsuccessful once dispersion forces are included.40
1.4.2 Present
Classical Simulation
Classical molecular dynamics simulations (MD) use point-particles to approximate atoms and molecules and an empirically derived interaction potential energy. Atomic/molecular motion is determined through integration of Newton’s Equations of Motion. I will discuss this theory more in section 2.2. They are the most widely used, as they are theoretically simple, easy to implement, and the oldest form of simulation. Perhaps the greatest benefit is the speed:
30 1.4. THEORIES AND SIMULATION OF ION SOLVATION the computational cost of MD simulations ranges from O(n2) to O(nlog(n)) if electrostatic approximations like Ewald summation are used. With these classical models it is possible to simulate approximately 100,000 atoms for 1 µs while for polarizable models 10,000 atoms can be simulated for 10 ns. For context, the timescale of protein diffusion is seconds, protein folding is ms, and ion channel gating is µs. MD has been successful in unraveling the biology of protein- protein interaction and conformational change. The field of allostery— the study of protein
(enzyme) binding has benefited as well. The conformational shift of Bacillus stearothermophilus lactate dehydrogenase, for example can be modelled in 50 ns of simulation.47 Drug Design and
"docking" of a compound to a macromolecule can be achieved through MD.
However, using Classical Mechanics to describe inherently quantum interactions has pre- dictable drawbacks. Past models do not include all of the interaction types and length scales that are needed to produce accurate results. The length scales of nanometer-scale materials behave drastically differently from what we experience in daily life. Although quantum mechan- ics tells us how molecules behave, we cannot intuitively understand the interactions of atoms and molecules at the nanometer scale. Designing proper experiments, theoretical models, and computational models to learn about these systems often suffers do to our lack of understand- ing. Even though the well-understood electromagnetic force is the fundamental force which generates atomic interactions, there are many types of forces under this umbrella which char- acterise interactions between different media. Interactions can be either short or long-ranged and their distinctions are less rigid at the nanoscale. In solvation, short-ranged interactions between ion and solvent are dominated by multipolar electrostatics (coulomb), induction, and exchange-repulsion. Long range-forces, which were overlooked in the past, are now proving to be important in solvation science48 49 as well as nanometer-scale device design, self-assembly and colloid science.48 Primary long-range forces include the Electrostatic (Coulombic), Polar
31 1.4. THEORIES AND SIMULATION OF ION SOLVATION
(acid-base), and Electrodynamic forces (van der Waals): Dispersion (London- instantaneously induced multipoles), Induction (or Debye) (permanent multipole and induced multipole), and
Keesom (permanent multipoles). Secondary long range forces are material specific, a sub- set of the above interactions, are macroscopic in nature, and include the hydrogen bonding, hydrophobic, hydration, and osmotic interactions, etc.48
There are many different ways to incorporate polarization. Point multipoles, fluctuating charge models, and harmonic spring models such as Drude oscillators are the main classical
MD options to model the above electrodynamic/electrostatic forces. Reference 50 modeled po- larization and charge transfer with a Drude oscillator model, and produced more accurate den- sities with temperature.50 In general, over-polarization is a major problem with many classical models, producing large errors.51 For example, modeling of DNA and RNA is largely inaccurate because the phosphate backbone is surrounded by ions in a highly polarized environment52 53
Understanding the folding process is dependant on understanding these interactions.
Non-electrostatic forces are also modeled with crude approximations which have varying degrees of success. Modeling of dispersion and exchange-repulsion is particularly rough.54 Van der Waals forces are modeled with a Lennard-Jones Potential55 which is treated as catch-all for corrections after simulation results are fit to a particular experimental property.
Ab Initio Simulation
Ab Initio Molecular Dynamics (AIMD) is the most realistic method of simulation because it uses the quantum mechanical Schrödinger equation instead of a crude empirical approximation for the interaction energy. Some basic theory will be discussed in section 2.3. Furthermore, these physically derived interaction potentials will be more transferable than classical models which are parametrized for specific conditions. It overcomes many of the drawbacks of classical
32 1.4. THEORIES AND SIMULATION OF ION SOLVATION
MD discussed above, as it is able to produce accurate condensed phase simulations capable of modeling electronic polarization effects and chemical reactivity56 (bond breaking/formation) that is sorely needed in problems of solvation science, among others. Some common methods include Born-Oppenheimer MD, Hartree-Fock MD, Kohn-Sham MD, Car-Parinello MD, and
Path Integral MD57
As ideal as this sounds, any simulation is restricted by the large computational cost of
AIMD— making applications of these methods very limited in size and time. An AIMD sim- ulation can only be done for approximately 100 atoms for 100 ps. This makes simulations of interfacial systems extremely difficult because of the number of atoms necessary. Electrostatic potentials, for example, must be determined in systems with a bulk-like region in order to find the convergence of the potential.
Neural Network Potentials
So far I have discussed the most common forms of molecular liquid simulation. Classical MD often represents a molecular charge distribution as a set of (often rigid) partial charges which approximate the multipole expansion, or as a set of point multipoles. While they are adequate approximations for some purposes, they fail to accurately represent the reality of the electron density. They are very computationally efficient— able to simulate the longest timescales of several nanoseconds. AIMD is far more accurate, but inefficient, as nuclei and electron density are treated quantum mechanically.
Recent advances in Neural Network Potentials (NNP) have created a new category of sim- ulation protocol where quantum accuracy can be achieved at the low cost of classical MD.
Neural networks are an advanced technique used to develop machine learning (ML) potentials.
Theoretically, these potentials must (1) express the analytic structure-energy relationship using
33 1.4. THEORIES AND SIMULATION OF ION SOLVATION machine learning, (2) use a first-principles training set of energies and forces, and (3) not con- tain any ad hoc assumptions.58 The general steps to construct an NNP are as follows. First, an initial set of electronic structure calculations is generated. This is often the most computation- ally expensive element which establishes the bottleneck. Next, the data needs to be transformed into appropriate input features that satisfy conditions of rotationally and translationally invari- ant system energy and permutation symmetry with the exchange of like atoms. Parameters are then varied to agree with the training data energies and forces in order to construct a Potential
Energy Surface. Finally the model is tested and simulations follow.
One such NNP is the DeePMD potential of Car and coworkers which has been shown to correctly represent the PES58 59 60 and produce accurate simulations of condensed phase systems.61 62 Metrics such as RDFs, packing parameters in water,63 organic liquids,59 and metal alloys64 have shown to very accurately reproduce AIMD reference data. In a subsequent study,
Zhang et al. 65 were able predict the dielectic response and the infrared spectra of liquid water, which was previously unattainable at the timescales of AIMD simulations. Drawbacks of NNP’s include many of the difficult aspects of neural networks in general. For example, the predictive power of the model is limited to the sample space (configuration space) on which it is trained.66
Additionally, reference data is obtained through the computationally expensive AIMD data we seek to avoid. Work to alleviate this issue includes active learning techniques to generate more sample space.64 Also, the NNP is essentially a black box: extracting physical principles from the simulation requires the thoughtful selection of the NNP and understanding of the input data, the training process, what it is trained for, and the output data.
Nonetheless, the benefits of NNPs outweigh the cost, and they represent a groundbreaking new tool in answering the outstanding questions of Solvation Science. Beyond the simple use of
NNPs for MD, they can be used for the calculation of various properties and incorporated into
34 1.5. INTERFACIAL POTENTIAL EFFECTS alternative models. There is further research needed to ensure that these models are physically grounded. In the context of molecular liquids, this includes whether or not the electrostatic forces are correct.
1.5 Interfacial Potential Effects
Motivation for the study of interfacial potentials and their effects is most directly linked to the establishment of a single-ion solvation free energy scale. Besides understanding the fundamental physics of interfacial potentials, they are essential for determining the thermodynamic scale that will help us to understand specific ion effects. First I will discuss this scale before introducing some of the more fundamental questions.
1.5.1 Thermodynamic Scale
Establishing the solvation free energy for an ion pair (salt) is simple because the total charge is zero. Thermodynamic and electrochemical experiments can only be applied to neutral macro- scopic systems. When the ion pair is separated, the ion has a charge and no single experiment can determine the solvation free energy. In order to determine single-ion solvation free energies, a conventional scale must be created, where everything is measured relative to the proton (H+) solvation free energy. Once, the proton value (or any single ion) is determined, all other ions can be determined from experimental bulk thermodynamic data.67 68 For a positive ion (P),
ex,conv ex ex ex ex µP = µP − µH+ = µP,b − µH+,b (1.6)
35 1.5. INTERFACIAL POTENTIAL EFFECTS and for a negative ion (N),
ex,conv ex ex ex ex µN = µN + µH+ = µN,b + µH+,b (1.7)
Different experiments have obtained highly accurate and agreeable results for both the pair free energies and conventional scale free energies,69 but not for the single-ion values.70 Beyond es- tablishing this scale relative to the proton, there must be an "extrathermodynamic assumption" that is used to rationalize the single-ion values.70 This is essentially an additional theoretical input that must be applied to deduce the single-ion quantities.
One commonly used assumption is the TATB hypothesis of Marcus71 — that two specific
(TA+/TB-) ions of opposite charge have the same absolute free energies in every solvent. This is because TA+ (tetra-phenyl arsonium) and TB- (tetra-phenyl borate) are large molecules of similar size, nearly spherical, and have an inert periphery.68 Because of this, no charge-specific solvent ordering will occur, and the solvation is nearly Born-like. The TATB hypothesis is similar to other methods such as those of Latimer72 73 where an interfacial-potential-free model is used to set the thermodynamic scale by partitioning bulk pair thermodynamic data into individual contributions.
Methods such as the Cluster-Pair Approximation (CPA), used by Tissandier,74 instead uses ion-water cluster data with an increasing number of waters (n) and bulk conventional scale ion hydration free energies to estimate the free energy to insert a proton into water. Free energies are calculated for small ion-water clusters relative to H+ and OH- in order to estimate the proton hydration free energy and set the scale. The original authors claimed this method is free of extrathermodynamic assumptions. As suggested by our lab75 and others76 77 78 79there is an extrathermodynamic assumption: that the conventional scale pair solvation free energies for small clusters converge to the bulk value as n increases. In other words, adding more water
36 1.5. INTERFACIAL POTENTIAL EFFECTS molecules to an anion cluster or cation cluster has the same binding energy, if they are large enough. In fact, there are effects from the net potential of water and hydration effects which contradict this assumption.
The other difficulty is that the real free energy in dragging an ion from gas phase to bulk liquid creates a contribution from the electrostatic potential of the vapor-liquid interface— the interfacial potential. I will discuss this contribution extensively throughout my thesis, particularly how it is created by the anisotropy of molecules at the liquid interface. Determining the value of this "phase potential" is essential for establishing this free energy scale, because it
ex ex differentiates the real free energy µi from the bulk value µb,i
ex ex µi = µb,i + qiφnp, (1.8)
Enthalpies and Entropies can be calculated similarly from temperature derivatives of the inter- facial potential. The real free energy can be experimentally measured from a combination of conventional ion hydration free energies, referenced to the proton, and cluster free energies of formation for ions.
Pollard and Beck80 asserted how the difference between the free energy values obtained with the Marcus and Latimer methods differed from the CPA method of Tissandier by approximately
+10kcal/mol. Noticing that the difference in simulation methods is largely due to one being interface-free and the other is not, they reasoned that the effective electrochemical surface potential of water should be -0.4 V. In another study,75 using classical and quantum methods, this value was found again by examining experimental results of ion distributions at interfaces and a re-analysis of CPA without any extrathermodynamic assumptions.
Shi and Beck81 compounded the confidence in this result by a quantum DFT simulation of bulk hydration free energy for Na+ and F- ions. By averaging the cation and anion bulk
37 1.5. INTERFACIAL POTENTIAL EFFECTS hydration free energy term that includes interfacial potential effects, the potential contribu- tion cancels and just the bulk value is found. Comparing the simulated bulk value with the experimentally obtained real free energy, results in an effective surface potential of -0.41 V.
Although evidence is accumulating for this liquid water effective surface potential value, we are still left with the question of it’s physical origin and value in other solvents. How does the anisotropy around an interface lead to this potential? Does the geometry of the interface affect this value?
Another important point which has been debated is the effect of a "local" interfacial potential due to anisotropic molecular arrangement around the ion. Pollard and Beck simulated TA+ and
TB- ion models in a surface potential-free system and showed the existence of a local interfacial potential. The free energies of these ions were similar, even in different solvents (water and dimethyl sulfoxide) after this local potential was removed. The existence and importance of the local potential cannot be overstated. It holds a key to understanding ion specific phenomena.
But first, I will discuss some subtle theoretical points of the interfacial potential.
1.5.2 Are they measurable?
The indeterminate nature of interfacial potentials goes back to the Gibbs-Guggenheim princi- ple which states that the electrostatic potential difference between two phases cannot be mea- sured.82 83 The principle further critisizes the use of classical electrostatics in electrochemical theories based on thermodynamic violations. Gibbs first wrote about the impossibility of mea- suring the potential difference between an electrolyte and electrode in 1889, and Guggenheim reformulated this in 1929 to say that the, "electric potential between two points in different media can never be measured and has not yet been defined in terms of physical realities. It is therefore a concept which has no physical significance.".82 He did state that the potential
38 1.5. INTERFACIAL POTENTIAL EFFECTS difference does have significance in electrostatics, where the electric field is an imaginary con- struct that determines the flow of electrons and ions, and their equilibrium is not static, but thermodynamic.
This debate over measurability has continued with various interpretations.83 84 85 86 Many widely used theories discussed in section 1.4.1 explicitly chose to ignore this principle by us- ing electrostatic theories in thermodynamic processes, which is nonphysical. For example, researchers still use theories like Debye-Hückel in biological studies of mammalian cells beyond the concentrations limits of the theory.83 Distributions of ions in solutions and colloid dispersion still make use of these theories.
Molecular dynamics methods, however, do not violate this principle and avoid the nonphys- ical assumptions of Poisson-Boltzmann theory and it’s derivatives. While there is no thermo- dynamic basis for electrostatic and non-electrostatic forces to be additive, it has not yet been conclusively disproven83 and remains the best method for investigation of interfacial potentials.
It is our aim that through a thoughtful combination of theoretical input and simulations, as well as comparison with experiment, that interfacial potential shifts felt by ions near interfaces can be determined.
There are three main reasons that make the study of interfacial potentials so difficult.
First, different experiments probe different physical effects.87 Each experiment probes a dif- ferent length scale, and each probe reacts differently to the potential gradients. For example, electrokinetic potentials, or the zeta potential ζ, are commonly measured by colloid scientists.
This is done by measuring the electrophoretic mobility of a colloid particle, and using the
Helmholtz-Smoluchowski law. It describes the electric potential at the plane where the mobile
fluid is separated from the attached fluid ("slipping plane"), and is not the same as the interfa- cial potential at two points deep in the phases, relevant to ion solvation. Second, as explained
39 1.5. INTERFACIAL POTENTIAL EFFECTS above, there is no direct way to measure the potential shift experienced by an ion crossing an interface.83 84 85 86 Third, the interfacial potential is not directly measurable. There is a slew of ambiguity in it’s definition accepted by experimental, theoretical, and simulation researchers.
It is of the utmost importance that a systematic definition across fields is developed. In the next section, I will discuss the definition further.
Even though the interfacial potentials are ambiguous and formally immeasurable, there are detectable physical consequences which support their existence and help in their determination.
One effect is the measurable ion distributions at interfaces caused by driving forces. In one experimental study, Conboy and Richmond88 investigated the interface of water and dichloro- ethane (DCE) and the distribution of TA+ and TB- ions. Using Total Internal Reflection
Second Harmonic Generation (TIR SHG), an optical technique, they found that ions were mainly located in the DCE phase, until an external potential of 0.4 V was applied, and the cation of the TA/TB pair nearest to the interface was replaced by the anion TB-. Beyond the confirmation that ion distributions at interfaces are affected by the interfacial potential, this result agrees with our value in section 1.5.1 obtained through analyzing thermodynamic data.
The TB anion has been shown to be more strongly hydrated than the TA cation.89 Another study by Faurudo et. al90 found that the anion TB moves towards hydrophobic surfaces, opposite to what is expected, and strongly suggests the influence of interfacial electrostatic driving forces. The force in this study agrees with the sign we predicted.
Another class of experiments are acid-base shifts which lead to the charging of neutral gas bubbles in water.91 92 93 Behaviour of carboxylic acids at the interface of water and other mediums is relevant to the transport of nutrients through cellular membranes, aerosol emissions from water and atmosphere, and origins of life. Creux et al91 studied the pH of the water surface when exposed to a low dielectric medium (oil, teflon, air) through electrophoresis and
40 1.5. INTERFACIAL POTENTIAL EFFECTS other methods, and found that anions were always drawn to the interface, creating a basic air/water interface. This is despite numerous MD results which showed the this interface is
93 acidic. Eugene et. al. exhibited the different pKa values for some of these acids near the water/vapor interface. The observed shifts are again consistent with an effective interfacial potential of -0.4 V.
Finally, there is the very recent discovery, by Zare’s group, that chemical reactions are spon- taneously induced by electric fields near the surface of a liquid.94 After accidentally finding that hydrogen peroxide (H2O2) was created in water microdroplets (micron-sized), they suggested the mechanism of hydroxyl (OH) radicals being created by OH- losing an electron. This fas- cinating result, would not be expected for inert, catalyst-free water, and suggests the role of an intrinsic interfacial potential at the surface of the droplet. This discovery showed a new method to produce hydrogen peroxide in an environmentally friendly way and a new possible method of disinfecting surfaces which could have profound impacts on society. Beyond that, it shows how an often overlooked physical quantity, the interfacial potential, can impact chemical synthesis, an area previously not considered.
This experimental evidence clearly shows that the pursuit of understanding and measuring interfacial potentials is not simply an intellectual exercise, but an important emerging field which can have a dramatic effects on a diverse range of fields— from green chemistry to the origins of life. However, the problem of studying this elusive quantity remains, and for that there must be a clear definitions between theory, experiment, and simulation.
1.5.3 What are we measuring?
In electrochemistry, potentials are often referred to as the Galvani (inner) potential φ and Volta
(outer) potential ψ. The Galvani potential is equal to the Volta potential plus the surface dipole
41 1.5. INTERFACIAL POTENTIAL EFFECTS contribution χ; φ = ψ + χ. The Galvani potential φ is the immeasurable quantity at the center of the material of interest. The Volta potential ψ, measurable in principle, is that outside of the condensed phase, close enough to pick up long-range electrostatic effects. The χ potential is often referred to as the "interfacial potential jump" going from one medium to another.87
Pratt argued that the separation of φ and ψ is not necessary after restricting the problem to conducting materials where there are non-existent macroscopic electric fields (and constant electrostatic potentials). Fluid dielectrics, although not conductors, can be determined to have macroscopic electric fields of zero.
Pratt’s work, which we follow closely, views the interfacial potential (or surface potential) as a contact potential as described by Laudau, Lifshitz, and Pitaevskii.95 The idea is introduced through the work done to remove a charged particle through the surface of a conductor— the work function. It must be a thermodynamically reversible process that depends on particle charge, thermodynamic state of the conductor, and the state of the surface itself. The relation
dφ of the work function to the charge density of the surface ρ(x) is given by the relation: dx =
R ∞ −4π −∞ ρdx. Another integration leads to,
Z ∞ φ(−∞) − φ(∞) = 4π xρdx (1.9) −∞
This relation shows that the difference in the work potential between the surfaces of a conductor is dependent on the dipole moments of the charges near the surfaces. The contact potential is the potential difference between the two conductors which prevents the flow of charges between them.95 From this it is clear that determining the contact (surface) potential is simply a matter of determining the the charge density. Further information can be gained through this formula and a substitution of the multipolar expansion of charge density as will be shown in sections
2.6-2.6.2. This theoretical definition provides a sensible way to determine this potential through
42 1.5. INTERFACIAL POTENTIAL EFFECTS simulation.
Landau et al. also discuss the chicken/egg aspect of the electrostatic potential difference in their discussion of the galvanic cell.95 The total contact potential in a closed circuit of conductors is nonzero, and called the electromotive force (e.m.f). They show that the total e.m.f of the cell is equal to both the sum of the interfacial potential shifts crossing all of the interfaces or as the differences of bulk ion chemical potentials between the phases (in electrodes and electrolytes). Despite this ambiguity, there are measurable physical effects which confirm the existence and physical importance of this potential.
Perhaps the trickiest part of understanding interfacial potentials beyond the theoretical definitions is that they differ depending on what experiment is used to measure them and what simulation is used to model them. In reality, the charge distribution around a molecule is composed of a dense nuclear core and a diffuse electron cloud. Ab initio simulation through
Density Function Theory approximates this distribution, while classical MD methods use a crude model of finite point charges. Clearly, the electrostatic potential measured with a test particle through a simulated media will differ based on the level of theory chosen as well as the nature of the test particle. Furthermore, an experimental comparison must account for the choice of probe.
In classical simulation, interfacial potentials are often calculated from the charge density.
The charge density is constructed from an ensemble average of the partial charges at spatial points through the entire system. In a spherical droplet system, this is accomplished through the the partial charge values and a radial distribution function (RDF) of all positions. Subsequent integrations using Maxwell’s Equation, ∇ · E = 4πρ, and the definition of electric potential,
E = −∇φ, results in electrostatic potential profiles through the liquid. Alternatively, the interfacial potential can be determined from constructing point dipole and quadrupole moment
43 1.5. INTERFACIAL POTENTIAL EFFECTS densities. Using these methods, the "probe" inherently penetrates into the molecular charge distribution. Similarly, in DFT calculations, the interfacial potential probe penetrates the interior of the molecule.
Examining interfacial potentials with these methods gives rise to an unexpected bulk po- tential value known as the Bethe Potential. Known under various terms as the "mean inner potential"96 or "exclusion potential",85 it describes the average electrostatic potential inside of a molecule. It is a transitionally invariant, bulk liquid property, which exists in nature. Both classical and ab initio methods contain the Bethe potential, but it is highly dependant on the
4π P R 3 2 2 theory level and model used. In general it is given by φB = − 6V i d rρ(r)r . The r term hints at it’s intrinsic relation to the quadrupolar terms in electrostatics. It tends to have a large positive value for quantum mechanical methods (∼ +3 V) and a small negative value for classical methods (∼ -0.5V).96 97 The bulk nature of the Bethe Potential leads to the un- surprising property that it cancels when crossing the two interfaces from vapor, to bulk, and then into a local cavity. Hence, it is common for a single surface potential estimate to contain this contribution, but a net potential containing two interfaces will not (no matter what theory level or model is used). Figure 1.2 shows a schematic of these interfaces for various systems.
Whether or not the Bethe potential is picked up in experimental studies is dependant on the probe used.98 85 99 96 If the probe is small and has high enough energy to penetrate the molecular center, there will be a contribution from the Bethe Potential. Experiments using high-energy electron diffraction or high-energy electron holography do penetrate into the molecular center.
These measurements should be most similar to the AIMD calculation of the surface (or net) potential of the solvent.98 100 In contrast, electrochemical experiments determine interfacial potentials by calculating the difference between the chemical potential of hydration of an ion and the real free energy change during transfer of an ion from vapor to liquid phase.101 102 103
44 1.6. SUMMARY
(a) (b) (c)
bulk solvent bulk solvent bulk solvent
Figure 1.2: A schematic representation of the possible simulated systems. (a) and (b) are slab systems, usually modeled with PBC Ewald boundary conditions. (c) is a droplet system with a cavity at the center. Both (b) and (c) contain two interfaces and thus contain the same Bethe potential cancellation, while (a) does not have this Bethe potential cancellation.
Here, the common probe is a proton, which does not penetrate the molecule due to the Van der Waals force. These experiments are most similar to classical MD, using a test particle as a probe, because the artificial Lennard-Jones potential exclusion prevents intrusion into the molecular center.
In summary, a sensible comparison between simulation and experimental interfacial poten- tial values is dependent upon understanding 1) Whether your experimental or simulation probe penetrates the molecular center (and picking up a Bethe potential contribution) and 2) how many interfaces are crossed, affecting your result (Bethe potential cancellation).
1.6 Summary
Many solvent models fail to produce accurate results for systems requiring physical effects be- yond their parametrization. While these models can perform for a small set of circumstances, they do not contain the proper physical theory to be transferable or reproduce experimental results for systems with slightly different specifications. Models that are capable of chemical and quantum physics can only be used for a limited number of systems due to their highly
45 1.6. SUMMARY computationally expensive nature. Ion solvation is one such problem which is critically impor- tant for a vast range of scientific areas. It is a complex problem where interpreting specific ion effects in experiment is difficult. Compounding on that, translating those observed effects and studying them with computational and theoretical models must be done in a methodical way which accounts for all underlying physical principles. A well-thought approach to studying ion solvation and the subtleties of the physics is essential. Only recently has the community begun to understand the profound effect that interfacial potentials can have on a range of physical phenomena, including the determination and unified scientific agreement on single-ion solvation
ex free energy values through the equation µ = µb + qφnp. Until recently, it was widely accepted that only a distant electrochemical surface potential φsp had an effect on free energies. Further, the electrochemical surface potential of water was determined from an averaging of various experiments, without regard for theory level or system details. This is clearly not the right ap- proach to measuring the interfacial potential shifts felt by ions near an interface. Recent work in our group provided further thermodynamic evidence for a water-vapor surface potential of -0.4
V and for the existence of a local potential φlp near an ion. But how is this potential generated by the molecules at the surface, and how does it differ from other non-aqueous solvents?
The problem of unraveling the relation between experiment, theory and simulation in Sol- vation Science is difficult for two reasons; 1) There are so many variables of simulation details, theory level, experimental methods, etc. in each of the three areas, that translation between them can be difficult and 2) Simulation models all suffer from the cost-accuracy problem which limits the calculation of properties like free energies and interfacial potentials with an accept- able accuracy. My thesis addresses the ion solvation problem by attempting to overcome these challenges with physically grounded analysis as follows:
• Chapters 3 & 4: What is the physical origin of φnp?
46 1.6. SUMMARY
– Using a simple model of classical water, eliminating variables such as boundary
conditions and complicated parametrization, how does this potential arise?
– Using a more complex organic molecule, are there general physical properties of the
potential shift which emerge? How do water and organic solvents differ in their ion
solvation properties?
• Chapter 5: Model development of ab initio accuracy at classical cost
– Can Neural Network Potentials produce electrostatic properties of liquid water?
47 Chapter 2
Theory
2.1 Preface
This chapter presents the theoretical background for the work which follows. The first section discusses classical MD and statistical mechanics. The next section provides a brief review of
DFT and AIMD techniques, focusing on those used in this thesis. Then I introduce the theory behind the Neural Network Potential methods that I use in my final project. Finally, I outline the thermodynamic techniques related to ion solvation, including the important contributions from interfacial potentials.
2.2 Classical Molecular Dynamics
Molecular Dynamics (MD) is a technique that uses the classical laws of motion to describe nuclear motion and can be used to calculate the equilibrium and transport properties of a many- body system.3 Like experiment, an individual atomic trajectory is not the sought after quantity, instead the average collective properties in time and space are determined. It is important to distinguish between time and space averaging, as they are important to the concept of ergodicity
48 2.2. CLASSICAL MOLECULAR DYNAMICS which is the underpinning of classical MD theory. For a short time simulation to produce real statistical mechanical results, it is necessary that the system satisfy the ergodic hypothesis3
ai(r) = hai(r)iNVE (2.1)
where a is some physical observable, the bar indicates a time average, and the brackets indicate an ensemble average. The ensemble average means an average over the phase space of the system for a given energy E. This hypothesis, which holds for the systems of interest here, says that the average of a over the time evolution of the system is equivalent to the ensemble average. Hence, a simple time evolution can be used to produce a full thermodynamic average.
Here I outline some of the tools and techniques used accomplish this.37
2.2.1 Equations of Motion
The most basic procedure is as follows. The first step in an MD system is to initialize with a set of atomic coordinates and velocities. Forces on each atom are computed and Newton’s equations of motion are integrated in order propagate a system trajectory. System averages can be calculated at each step. The Verlet method104 is one of the most popular methods of integration, with each new position is evaluated by,
f(t) r(t + δt) = 2r(t) − r(t − δt) + δt2 δt2 (2.2) m and the velocities constructed with,
r(t + δt) − r(t − δt) v(t) = (2.3) 2δt
49 2.2. CLASSICAL MOLECULAR DYNAMICS
The kinetic energy and system temperature can immediately be constructed from this knowl- edge of system coordinates and velocities at the microscopic level and can be translated to the macroscopic level through statistical mechanics. Once the thermodynamic state is speci-
fied, thermodynamic properties such as density, chemical potential, heat capacity, structural quantities, and time correlations can be determined.
2.2.2 Thermodynamic Ensembles and Equilibration
Equilibration of the system ensures that a statistical mechanical ensemble is maintained. The
Microcanonical, NVE, Canonical, NVT, and Isothermal-isobaric, NPT, are commonly used.
Constant Volume, V , and number of particles, N, are easily achieved in simulation. Constant temperature is maintained through use of a thermostat such as Nose-Hoover or Berendsen.
The Nose-Hoover thermostat’s instantaneous kinetic energy (and velocities) fluctuate while temperature is held constant due to introduction of a heat bath to the Hamiltonian. Constant pressure can be achieved with a barostat, which scales unit-cell vectors in response to the stress tensor.
In practical applications, one or more of these ensembles may be used depending what system is being modeled. Simulation of interfaces between different phases or chemical species may involve separate equilibration procedures before combining them.
2.2.3 Classical Force Fields
Perhaps the most important choice in MD simulation is deciding how to model the poten- tial energy. Commonly referred to as the force field, it determines the functional form and parametrization of the atomic/molecular interaction energy. The potential energy function is usually divided into bonded and non-bonded interactions. The former typically meaning inter-
50 2.2. CLASSICAL MOLECULAR DYNAMICS molecular (chemical) bonds between atoms and the later meaning intramolecular interaction.
Bonded terms can be simplified through use of "rigid" models in which molecules are seen as ball-and-stick entities where molecular vibrations can be ignored. Alternatively, harmonic springs approximate the atomic bonds, out-of-plane angles, and torsion of molecular motion.
Non-bonded interactions model the physical forces of dispersion, repulsion, and electrostat- ics. The commonly used Lennard-Jones potential uses a repulsive r−12 term to imitate the effects of the Pauli exclusion principle of overlapping electronic orbitals, and an r−6 term for the attractive, long-range dispersion (van der Walls force).7 In the Lennard-Jones potential shown below, parameter has units of energy, describing the depth of the potential well, and
σ has units of distance and describes the distance at which the potential is zero.
"σ 12 σ 6# U = 4 − (2.4) LJ r r
Each atomic species is assigned an i and σi with interaction between species determined by mixing rules. Electrostatic interactions are determined by a simple Coulomb interaction energy,
X qiqj Ues = (2.5) i6=j rij
where charges qi are the partial charges of each atom.
In practice the parameters and partial charges above are determined either by comparison with quantum chemical calculation or through experimental properties. Simulations are run with initial guesses to calculate properties of solids, liquids, and gases, and then compared with reference values. Parameters like i and σi are adjusted until satisfactory agreement. Because of this approach, classical force fields are only successful in the subset of systems for which they are parametrized. As a result they are non-transferable.
51 2.2. CLASSICAL MOLECULAR DYNAMICS
2.2.4 SPC/E Force Field
Modeling water is integral to many bio-molecular, condensed matter, and chemical systems, leading to intense efforts in force field development. The detailed solvent interactions are critically important and are not correctly described by continuum models. Classical water models generally consist of three to five atomic sites with various partial charge values and locations with rigid or non-rigid Lennard-Jones parameters. The SPC/E (simple-point charge, extended) water model105 106 is popular for it’s simplicity and computational efficiency. It consists of 3 atomic sites: oxygen and two hydrogen atoms, with a partial charge on each site. SPC/E refined the SPC model by adding a term to correct for average polarization which improves the density and diffusion constant. The SPC/E model uses partial charges qoxygen = −0.8476 and qhydrogen = +0.4238 with the equilibrium bond length of 1.0 Å and an
H-O-H angle of 109.47◦.
The SPC/E model inter-atomic potential is given by the force field,
1 h i 1 U = k (r − req )2 + (r − req )2 + k (θ − θeq )2 + U + U (2.6) SP C/E 2 b OH1 OH OH2 OH 2 a HOH HOH LJ ES
where rOH are bond lengths and θHOH is the H-O-H angle. Here, we use the rigid model, where spring constants ka and kb are very large. For the purposes of theory, this model is ideal as it is one of the most basic models, yet it performs remarkably well in comparison with a wide range of experimental properties.107 108 109
52 2.2. CLASSICAL MOLECULAR DYNAMICS
2.2.5 OPLS-AA
The OPLS-AA force field was parametrized for experimental properties of liquids. It is non-rigid and the functional form of the potential is,110
X 2 X 2 X UOP LS−AA = Kr(r − r0) + kθ(θ − θ0) + f(dihedrals) + ULJ + UES (2.7) bonds angles dihedrals
q q where the Lennard-Jones potential uses mixing rules Aij = AiiAjj and Cij = CiiCjj.
2.2.6 Boundary Conditions
The boundaries of the simulation region depend on the system being modeled. In the case of a finite system, a bounding potential is often used. In my droplet study, I use a simple
1 2 half-harmonic bounding potential wall U = 2 k(r − r0) to ensure that molecules do not leave the simulation box. If the system is properly equilibrated, the number of molecules interacting with this potential is minimal.
In the case of extended systems, such as bulk liquids or flat interfaces, a more complex treatment is needed using periodic boundary conditions (PBC). The number of molecules that would be needed for these large systems is computationally unfeasible, so a relatively small system box is replicated in each dimension with conditions on each boundary. The Ewald potential is commonly used to implement PBC. For a box of size V = L3 both the Coulomb interaction potential and the energy between an ion and it’s periodic images is expressed through
Ewald summation: k2 4η2 X erfc(η|r + nL|) X 4π e ikr ψewald = + 3 2 e (2.8) n |r + nL| k6=0 L k
η is the convergence parameter, n is the real-space lattice vector, and k is the Fourier-space
53 2.3. AB INITIO MOLECULAR DYNAMICS lattice vector. The electrostatic potential energy is then:
N N N X X ξ X 2 U = qiqjψewald(rij) + qi (2.9) i=1 i=1 2L i=1
The last term in this equation is the self-energy correction term, which must be evaluated for the specific system geometry (ξ = −2.837297 for cubic geometry). This correction arises out of the need for a neutralizing background and accounting for the image charge effects.1117
While this method drastically improves computational efficiency, there are some issues.
This artificial system produces artifacts outside of physical reality.112 113 111 Correction terms for solvent polarization, finite-size effects, thermodynamic single-ion solvation properties (enthalpy, entropy, heat capacity, etc.) have to be included.112
2.3 Ab Initio Molecular Dynamics
Highly accurate ab initio molecular dynamics (AIMD) methods create finite-temperature dy- namic trajectories by performing electronic structure calculations during the simulation. Ap- proximate solutions to the Schödinger Equation are achieved through Density Functional The- ory (DFT). For a system of N nuclei and Ne electrons, the total Hamiltonian is given by
H = Te + Vee + VeN + TN + VNN (2.10)
In classical MD, this Hamiltonian is reduced to a force field approximation as in sections 2.2.3-
2.2.5. In DFT, the Hamiltonian is slightly simplified with the Born-Oppenheimer Approxima-
114 tion, which allows for the electronic and nuclear coordinates to be separated. That way, VNN is a constant and TN is equal to zero. There are many AIMD methods and codes available
54 2.3. AB INITIO MOLECULAR DYNAMICS to researchers. I will narrow the scope to the methods of Density Function Theory115 (DFT) and the Gaussian Plane Wave approach (GPW)116 117 used in the open source software package
CP2K.118
According to DFT, the electronic ground state structure is determined through the electronic density distributions n(r) and ground state energy E. This is based on the Hohenberg-Kohn
119 Theorems which state that (1) The external potential Vext and energy E are functions of density n(~r) and (2) The exact density n(~r) is that which minimizes the total energy. So, there is an exact one-to-one correspondence between ground-state electron density and the wave function and between the wave function and the many-body Hamiltonian. The Kohn-Sham ansatz115 56 makes the problem solvable by replacing the interaction electron potential with an effective non-interacting electron potential Veff . The reduced Hamiltonian is then:
H = T [n] + V [n] + U[n] = T [n] + Veff [n] (2.11)
R e2 0 where Veff [n] = V (~r)+ |~r−~r0| dr +Vxc[n] and Vxc[n] is a chosen exchange-correlation functional.
The accuracy of the electronic structure calculation depends on a good approximation to Vxc.
This reduced problem leads to the Kohn-Sham Equations which must be solved self-consistently.
The Gaussian Plane Wave method uses atom-centered Gaussian orbitals to represent the wave functions (the Kohn-Sham matrix) and auxiliary plane waves to represent the electronic density.
The energy is efficiently solved for with Fast Fourier Transform Methods (FFT).118
Once the ground state energy is determined, the dynamics are solved for using Born-
Oppenheimer Molecular Dynamics.120 The potential energy surface is the ground state energy and the atomic coordinates are treated classically and propagated using Newton’s Equations.
Like classical MD, it is up to the researcher to choose suitable simulation-specific functions: variables of AIMD include the Gaussian basis sets for the wavefuntions, exchange-correlation
55 2.4. NEURAL NETWORK POTENTIALS functionals, and pseudopotentials to describe the nuclei. Determination of electrostatic prop- erties such as the multipole moments of a charge distribution are achieved with maximally localized Wannier functions (MLWFs).121 In AIMD, the charge is diffuse and must be assigned to atoms in a chemically reasonable manner.122 I’ll discuss this further in section Chapter 5.
2.4 Neural Network Potentials
The goal of Neural Network Potentials (NNP) is to reduce the electronic structure problem of
DFT/ AIMD approaches with a Machine Learning derived potential. Instead of a complicated analytical Potential Energy Surface (PES), NNP’s develop a neural network structure-energy relationship. This is physically correct, as the Born-Oppenheimer Approximation states45 that the Hamiltonian itself is completely defined by the atomic positions, nuclear charges, and total charge of the system.58
This approach arose out of the technique123 124 125 of representing the PES through a highly general, flexible, and adaptable set of mathematical functions which are optimized to fit ab initio data instead of empirical results.58 These "mathematical potentials" were developed in the 1990’s and use a large number of terms. Machine Learning methods such as artificial neural networks, kernel ridge regression, and support vector machines have been used to efficiently and accurately construct the PES.126 127 128
Neural networks (NN) are theoretically able to produce any multi-dimensional function.129
In brief, neural networks are computational processing systems modeled after the human brain, and are composed of many simple, interconnected processing units, nodes, which make decisions based on some response to external input.130 These nodes are organized into layers of varying size. NNs with many hidden layers are referred to as "Deep" NNs. Here, I use a fully-connected feed-forward NN, meaning each node is connected to all other nodes in the neighboring layers
56 2.4. NEURAL NETWORK POTENTIALS
Figure 2.1: Schematic of a neural network with 2 hidden layers. Input layer values are feed- forward to the output layer. In the case of a NNP, this output layer is total energy E which is the sum of individual atomic energies. Inspired by Ref. 130. with a forward flow of outputs.
j kl Important elements of NNs are the bias weight bi , weight parameters Wij , and activation
j function φi as shown in Figure 2.1 . The learning, or training, process is similar to the traditional sense of the word. Training data is fed to the first layer with a linear or nonlinear transformation
j known as the activation function. The value of node yi is then given by
Nj−1 j j j X j−1,i j−1 yi = φi bi + Wk,i · yk (2.12) k=1
The bias and weight parameters are free parameters to be optimized. After one iteration of the full training set through the NN, called an epoch, the output is optimized through a process of minimization of the loss function (or objective function). The number of iterations in this
Batch Normalization procedure is variable, and weights and biases are only updated after one epoch. The loss function L is usually the mean squared error of the quantity of interest. The
57 2.4. NEURAL NETWORK POTENTIALS errors are then back-propagated through each step starting from the final parameter set at t, similar to a steepest descent algorithm,
∂L W (t + 1) = W (t) − r · (2.13) ∂W (t)
Here r is the learning rate, which affects convergence.
Turning to the specific case of NNPs used in atomic simulations, I’ll restrict my discussion to an overview of the procedure of Wang et al. 63 contained within the open source package
DeePMD-kit. An extremely important element of NNs is processing the input data. The in- put data are the atomic coordinates along with the energies that are used for optimization
(learning), and the output is the inter-atomic potential energy which has permutational, ro- tational, and translational symmetry.131 The training data consists of an ab initio trajectory of coordinates and calculated energies, forces, and virial for the system. Coordinates for each frame of the trajectory are then processed into a symmetry-preserving set of coordinates. For a system of N atoms, with lab coordinates {R1, R2, ..., RN }, neighboring atoms are determined by Rij = Ri − Rj within a certain cutoff around the atom in order to preserve translational symmetry. Atoms are chosen from this list by a user-defined rule to form a local frame. The coordinates are then rotated into this local frame to preserve rotational symmetry and finally are sorted according to the chemical species s(i), and then ascending inverse distance to atom i to preserve permutational symmetry. These descriptors Dij are produced for each coordinate in the local frame, with the option of full-coordinate or only radial coordinate information. The full information is given by:59
1 xij yij zij {Dij} = { , , , } (2.14) Rij Rij Rij Rij
58 2.5. THERMODYNAMICS OF ION SOLVATION
These descriptors are fed through the deep neural network NW (i) that determines Ei, the
α energy of atom i, through it’s coordinate descriptor: Ei = NW (i)({Dij} }j∈N (i)) where N (i) is the neighbor list. In this network, the activation function φ is the nonlinear hyperbolic tangent of the linear combination of weights and biases as described above. The loss function
L(p, pf , pξ) given by:
p p p L(p , p , p ) = ∆E2 + f Σ ∆|F |2 + ξ ∆||Ξ||2 (2.15) f ξ N 3N i i 9N
where ∆E, ∆Fi and ∆Ξ are the root mean squared error in the energy, forces, and virial, respectively. Forces on the atoms and virial are computed by gradients of system energy with respect to position. Optimization is done using the Adam stochastic gradient descent method.132 Training data is divided into batches with loss function and gradient computed at each training step only from data within the current batch. A more detailed discussion of the network can be found in Ref 63. Some network parameters are user defined and can be adjusted to ensure an adequately converged network by analyzing the error in the training and testing data as the batch iterations proceeds. Once this is achieved, the biases and weights for the trained network are frozen, and it can be used on a new system of input coordinates.
2.5 Thermodynamics of Ion Solvation
Reprinted (adapted) with permission from
Doyle, Carrie C., Yu Shi, and Thomas L. Beck. "The importance of the water molecular quadrupole for estimating interfacial potential shifts acting on ions near the liquid–vapor in- terface." The Journal of Physical Chemistry B 123.15 (2019): 3348-3358. Copyright 2020.
American Chemical Society.
59 2.5. THERMODYNAMICS OF ION SOLVATION
For ion solvation science, one of the most important quantites to determine is the chemical potential, or partial molar Gibbs Free Energy
δG ! µi = |T,P,n (2.16) δni
Free energies tell us about the phenomena of binding, equilibrium constants, rate constants activity coefficients, osmotic coefficients, surface tension, and phase transitions. In the following section I will discuss the method we use to calculate free energies, Quasichemical Theory (QCT), which uses a spatial partitioning of the thermodynamic process.
2.5.1 The Potential Distribution Theorem
The chemical potential, µ, of an ion in dilute solution is given by
βµ = ln(ρΛ3) + βµex (2.17)
1 where β = kT , ρ is the density, and λ is the thermal de Broglie wavelength. The excess chemical potential µex or absolute solvation free energy is the difference between the chemical potential of species i and an ideal gas under the same conditions.3 There are several ways to calculate excess chemical potentials from simulation, one method is the Widom Particle Insertion Method133 used in the Potential Distribution Theorem.97 111
The Potential Distribution Theorem expression for the excess chemical potential of a molecule or ion is111 97
ex βµ = lnhexp(β∆U)i = lnhexp(−β∆U)i0 (2.18)
where ∆U = UN+S − UN − US is the interaction energy of the molecule or ion (N) with the solvent (S). The sampling in the first expression, h...i, includes the solute fully coupled to the
60 2.5. THERMODYNAMICS OF ION SOLVATION
solvent and the sampling in the last expression, h...i0, involves the solvent and solute separately with no coupling.97
These expressions in equation 2.18 can be rearranged to give
ex Z βµ β∆Ui βε e i = he i= Pi(ε)e dε (2.19)
for the coupled case and
ex Z (0) −βµi −β∆Ui −βε e = he i0 = Pi (ε)e dε (2.20)
(0) for the uncoupled case where Pi(ε) = hδ(ε − ∆U)i and Pi (ε) = hδ(ε − ∆U)i0 are the prob- ability densities of ion/solvent interaction energies for a coupled and an uncoupled trajectory, respectively. If the distributions overlap, the energy at which these two distributions cross gives the exact excess chemical potential. If they do not overlap, we use an alternate method known as Quasichemical Theory. This method arose out of the fact that the tails of the distributions are poorly sampled in an MD simulation. Spatial partitioning of the thermodynamic process can yield near Gaussian distributions Pi(ε)
2.5.2 Quasichemical Theory
Using this spatial partitioning and a repulsive wall potential M the above excess chemical expression 2.18 can be rewritten as,
ex −βM(r) −βM(r) −β∆U βµ = lnhe i − lnhe i0 − lnhe iM
−β∆U = ln x0(r) − ln p0(r) − lnhe iM (2.21)
ex ex ex =µis (r) + µpk(r) + µos (r)
61 2.5. THERMODYNAMICS OF ION SOLVATION
This equation can be derived from a thoughtful inclusion of an expression equal to unity. These terms are referred to as the inner shell (is), packing(pk), and outershell (os) terms. Inner shell term xo is the probability of finding no solvent in the cavity formed by the repulsive wall with full ion-solvent interaction. The packing term po is the probability of finding no solvent in the cavity of the ion-free solvent. The outershell (or long-ranged) term describes the remaining interaction of ion with solvent beyond the excluded volume.
The steps of this thermodynamic cycle are
1) open a cavity larger than the ion;
2) insert a point charge at the cavity center;
3) convert the point charge into the real ion;
4) remove the cavity constraint and allow the solvent molecules to move into direct contact
with the ion.
The corresponding free energy changes are:
1) the cavity formation free energy (or packing);
2) a purely electrostatic part of the electrochemical potential due to insertion of a test
charge into the cavity;
3) a small quantity that includes a dispersion contribution depending on cavity size;
4) the inner-shell or chemical part of the free energy that includes all chemically specific
effects due to direct contact of the ion with the solvent, such as electrostatics, polarization,
dispersion, and charge transfer.
62 2.5. THERMODYNAMICS OF ION SOLVATION
The textbook expression for the electrochemical potential134,135 for an ion i is
ex ex µi = µb,i + qiφnp, (2.22)
ex where µb,i is the bulk hydration free energy and qi is the ion charge. If we link this expression to the QCT partitioning, it is sensible to define φnp as the purely electrostatic net potential at the center of a cavity into which the ion is inserted [step (2) above]. Interpreting what this net potential means depends on the system being studied. Accurate determination and definition of this net potential is essential to establishing a single-ion solvation free energy scale.
2.5.3 Interfacial Potentials
As discussed above, an interfacial potential is the electrostatic potential produced upon moving a charge across regions of nonuniform charge, such as those at phase boundaries or other interfaces. As shown in equation 2.22, if the charge is nonzero, there will be a contribution to the excess chemical potential. As discussed in section 1.5.2, the measurability of these potentials has been called into question for many years, but recent work has shown that there are measurable physical effects. This is predicated on a careful consideration of experiment and simulation specifics.
It is critically important that the researcher understands the boundary conditions involved and their bearing on the interfaces. If periodic boundary conditions are used, there is no explicit surface interface that separates the bulk liquid from the vapor. The real hydration free energy is then
ex ex µ = µint + qφsp, (2.23)
where µint is the intrinsic free energy evaluated without that surface interface and φsp is the
63 2.5. THERMODYNAMICS OF ION SOLVATION
Liquid
Vapor
Figure 2.2: A schematic representation of the studied system - a spherical liquid water droplet
(blue) containing a spherical cavity (grey). Potentials φsp and φlp denote the potential shift due to crossing the liquid-vapor interface of radius Rdrop, and the cavity-liquid interface of radius
λ. The net potential at the cavity center is then: φnp = φlp + φsp interfacial potential across that interface. Alternatively, it can be defined in terms of the net
ex potential φnp the potential at the center of a neutral cavity and the bulk free energy µb as follows:
ex ex µ = µb + qφnp, (2.24) which is equivalent to 2.22 above. This definition is appropriate for simulations of a finite droplet or a PBC slab, which both contain two interfaces. In those cases, the two “interfaces" that are crossed upon moving from the outer vapor phase into the cavity are: the liquid-vapor interface (yielding the surface potential φsp) and the cavity-liquid interface near the ion (yielding the local potential φlp). A schematic depicting these potentials is provided in Figure 2.2. This definition produces:
φnp = φlp + φsp (2.25)
A second motivation for the physical significance of the net potential is the exact QCT
64 2.6. MACROSCOPIC INTERFACIAL ELECTROSTATICS expression for the ion excess chemical potential derived in Ref. 136:
ex (0) n ex ex µ = −kBT ln Kn ρW + kBT ln p(n) + µXWn − nµW (2.26)
(0) where Kn is the equilibrium constant for binding n solvent molecules to the ion in the gas phase, ρW is the solvent number density in the liquid, p(n) is the probability to observe n solvent
ex molecules bound to the ion in the liquid, µXWn is the free energy to insert the ion/(n-solvent)
ex cluster, and µW is the free energy to insert a single solvent molecule into the bulk liquid. In numerical implementations of QCT, n typically includes the first solvation shell or roughly 4–8 water molecules.111 Thus the cavity size for inserting the cluster will be on the order of 4-6 Å.
ex The net potential can be seen to be located in the µXWn term in the free energy. If a cavity size (for inserting the cluster) is chosen so the inner shell and packing parts of that free
ex energy largely cancel, then the predominant contribution to µXWn will be a long-ranged (largely electrostatic) free energy that is close to the net potential we have defined above. This cavity length scale occurs roughly at the peak of the radial distribution function between the cluster and the next layer of solvating molecules.
2.6 Macroscopic Interfacial Electrostatics
The net potential φnp has physical significance which cannot be overlooked in ion solvation thermodynamics. In the present section I outline the theoretical methods used to investigate it’s origin. In particular, I analyze the molecular multipole distributions at each interface and how they generate the net potential at cavity center. This discussion relies on Refs. 137, 138,
139 and 140. Reprinted (adapted) with permission from
Doyle, Carrie C., Yu Shi, and Thomas L. Beck. "The importance of the water molecular
65 2.6. MACROSCOPIC INTERFACIAL ELECTROSTATICS quadrupole for estimating interfacial potential shifts acting on ions near the liquid–vapor in- terface." The Journal of Physical Chemistry B 123.15 (2019): 3348-3358. Copyright 2020.
American Chemical Society.
2.6.1 Multipole Expansion of a Charge Distribution
Multipole moment expansions can be used to approximate a charge distribution and the result- ing potential. Following Jackson’s derivation138 of the macroscopic electrostatic equations from the microscopic definitions, we consider molecules as collections of bound charge. Performing a spatial average leads to the picture of averaged microscopic charge density as the sum of molecular multipole moment contributions.
For an individual molecule n at position xn with charges j at positions xj = xn + xjn the
first three molecular multipole moments (charge, dipole, and quadrupole) are given by:
X qn = qj j(n)
X pn = qjxjn (2.27) j(n)
X (Qn)αβ =3 qj(xjn)α(xjn)β j(n)
The microscopic average charge density hηi can then be written as the sum of the macro- scopic multipole densities— ρ(x), P(x),Qαβ(x) — which are obtained by summing over molecules.
Additional summation of these multipole moments over the moment order, we obtain the aver- aged microscopic charge density as the sum of the macroscopic multipole densities ρ(x), P(x),
138 and Qαβ(x):
2 X ∂ hη(x, t)i = ρ(x, t) − ∇ · P(x, t) + Qαβ(x, t) + ... (2.28) αβ ∂xα∂xβ
66 2.6. MACROSCOPIC INTERFACIAL ELECTROSTATICS
Substituting eq (2.28) for the charge density ρ and performing the appropriate integrations of
Possion’s Equation,95,141
∇2φ = −∇ · E = −4πρ (2.29) we can determine individual multipole interfacial potential contributions.84,142
2.6.2 Coordinate Systems: Cartesian and Spherical
Wilson, Pohorille, and Pratt142 showed that, for a planar interface, there are contributions to the interfacial potential from both dipole and quadrupole moment densities. In Cartesian coordinates, using (2.28) for ρ(z0) and integrating
Z z δφ(z) = 4π z0ρ(z0)dz0 (2.30) zv from vapor to liquid, the potential across a planar interface is given by:
Z zl 0 0 δφ = 4π Pz(z )dz − 4π[Qzz(zl) − Qzz(zv)] (2.31) zv
Higher order multipoles do not contribute because they involve spatial derivatives of higher- order terms. Such gradients do not exist in the bulk fluid, showing that, for the planar surface, the expansion truncates after the quadrupolar term. This relationship has been demonstrated throughout the literature for different water models.98,142–144
Horvath et al.143 later derived the expression analogous to eq (2.31) for a spherical coordi- nate system. Note that the unprimed quadrupole density Q is in cartesian coordinates, and the
0 D Q primed Qrr is in spherical coordinates. With δφr(r) = δφr (r) + δφr (r) the dipole contribution
67 2.6. MACROSCOPIC INTERFACIAL ELECTROSTATICS is given by Z r D 0 0 δφr (r) = 4π Pr(r )dr (2.32) 0
Q and the quadrupole contribution δφr (r) is given by
Q Q,1 Q,2 δφr (r) = δφr (r) + δφr (r) (2.33)
where
Q,1 0 δφr (r) = −4πQrr(r) (2.34) and
Z r 0 Q,2 dr 0 0 0 δφr (r) = 4π [TrQ(r ) − 3Qrr(r )] (2.35) 0 r0
The planar geometry and spherical geometry both contain multipolar elements, though the contributions from each multipolar element are take a different functional form.143 For the planar and spherical geometries, the dipole contributions to the interfacial potentials are of the same form, reflecting their dependence on water dipole orientation (polarization) alone.
For a spherical system, the quadrupole contribution has two terms (eq 2.33). The first term
Q,1 δφr (r) is a bulk property of the liquid, which has a form analogous to that for the quadrupole part of the planar case above (eq 2.31). In either geometry, it cancels when crossing the two
“interfaces" - passing from vapor, through liquid, and into cavity center - and is equivalent to the Bethe Potential96 described in section 1.5.3
4π φ = ρTrQ (2.36) B 6 0
where ρ is the molecular liquid density and Q0 is the primitive (traced) molecular quadrupole
68 2.6. MACROSCOPIC INTERFACIAL ELECTROSTATICS moment.98,143,145,146 Because typical ions do not penetrate into the solvent molecular interior, this contribution is not important for ions near interfaces.96
Q,2 143 The second term δφr is created by the spherical symmetry breaking of an interface.
For the droplet system, this residual quadrupole contribution is due to the curvature of the separate cavity-liquid and liquid-vapor interfaces. The sign of this term is determined by the quadrupole density profiles. It has the same sign for both interfaces, and for the liquid-vapor interface this second term approaches zero in the limit of infinite droplet radius.
We emphasize that, in the above equations (2.28)-(2.36), the traceless form of the multipole moments cannot be used unless the potential is calculated outside of the charge distribution.
The relationships above follow from Jackson’s expansion of the microscopic charge density into molecular multipole moments. He showed that when the partial charges are viewed as a molecule, the quadrupolar contribution to the microscopic charge density is traceless. This differs from the quadrupolar contribution contained in the macroscopic Maxwell equations, where the trace remains. This can be explained by noticing that the quadrupolar trace terms are not eliminated completely, but move to the macroscopic charge density, or first moment, when outside of the distribution.98,138 When considering the electrostatic potential at a point, the quadrupolar trace term must be considered while within the charge distribution but can be eliminated outside of it where the macroscopic charge density term, and hence the quadrupolar trace is zero.98,140,146 This allows us to use the following direct method for calculating the interfacial contributions to the net potential at the cavity center (or any zero charge density region).
69 2.6. MACROSCOPIC INTERFACIAL ELECTROSTATICS
2.6.3 Molecular Multipole Expansion of Electrostatic Potential
The electrostatic potential at x = 0 due to a molecule n at position xn with charges j at
138 positions xj = xn + xjn is given by:
" # X 1 1 1 1 φ(0) = qj − (xjn)α∇α + (xjn)α(xjn)β∇α∇β − ... (2.37) j xn xn 2 xn
The tensor notation indicates a sum over cartesian coordinates. The origin of our multipole calculation is the water oxygen. The moments are shown in eq (2.27) except the quadrupole
3 P 1 2 moment is replaced by the traceless form, (θn)αβ = 2 j qj[(xjn)α(xjn)β − 3 xjnδαβ]. Outside of a charge distibution, the potential is then,140
" 1 1 1 1 # φ(0) = qn − (pn)α∇α + (θn)αβ∇α∇β − ... (2.38) xn xn 3 xn
Summing eq (2.38) over n molecules and taking an ensemble average produces an estimate of the electrostatic potential in terms of the charges, charge coordinates and molecular coordinates.
With the cavity center located at x = 0, this provides an approximation to the net potential
φnp. (See reference 140 for the octupole term.) Equation (2.38) can be rewritten in more compact notation as 1 φ(0) = Tq − T (p ) + T (θ ) (2.39) n α n α 3 αβ n αβ where T is the molecular coordinate tensor:140
1 T = xn 1 (x ) n α (2.40) Tα =∇α = − 3 xn xn 2 1 3(xn)α(xn)β − xnδαβ Tαβ =∇α∇β = 5 xn xn
70 Chapter 3
Water Liquid-Vapor Interfacial
Potential Shifts
3.1 Preface
Reprinted (adapted) with permission from,
Doyle, Carrie C., Yu Shi, and Thomas L. Beck. "The importance of the water molecular quadrupole for estimating interfacial potential shifts acting on ions near the liquid–vapor in- terface." The Journal of Physical Chemistry B 123.15 (2019): 3348-3358. Copyright 2020.
American Chemical Society.
In this Chapter I present the results from my water droplet study. I attempt to answer the question that I posed in section 1.6: What is the physical origin of φnp? The interfacial potential shifts that act on ions near the water liquid-vapor interface are essential for a proper understanding and determination of the single-ion solvation free energy values. Our work builds on previous related analyses in Refs 97,98,143,145–147. I investigate the radial size-dependence of the potential from both the cavity-liquid and liquid-vapor interface. The question we address
71 3.1. PREFACE is: how do the molecular multipoles contribute to the observed average net potential at the cavity center? We choose droplets large enough to exhibit quasi-bulk behavior in the region between the cavity and the surface. For that bulk region, the average charge density is zero, so there is negligible contribution to the net potential. I simulate droplets consisting of 250–1500 water molecules with cavities of variable size embedded at the droplet center. By modeling droplets, all electrostatic interactions through direct summation of the Coulomb potential are included.
Thus, we avoid periodic boundary finite size effects due to the Ewald treatment of long-ranged electrostatics.113,148,149 Recently, droplet models have been examined in relation to other aspects of bulk behavior.150,151 The range of droplet sizes modeled here was chosen to assess the distance dependence of the multipole contributions from the liquid-vapor surface. Here we utilize only the SPC/E water model to illustrate the basic physics of the origin of the net potential.
The distance dependence of the various multipole net potential contributions can be an- alyzed by a simple scaling expression in the spherical droplet configuration (with the cavity located at the droplet center):
Z R+∆R 1 2 φ(0)l,np ∼ ρMl 4πR dR (3.1) R Rl+1
where the dipole contribution is l = 1, quadrupole l = 2, etc., and ρMl is the multipole density of order l in the inhomogeneous region. The dipole contribution is special because it does not decay in magnitude with distance from the cavity center (see also Ref. 146).
Next, consider the quadrupole contribution. Assuming the quadrupole density is roughly constant over the inhomogeneous domain of size ∆R (and that ∆R << R):
Z R+∆R 1 2 φ(0)l=2,np ∼ ρM2 4πR dR ∼ ρM2 ln(1 + ∆R/R) ∼ ρM2 ∆R/R (3.2) R R3
72 3.2. COMPUTATIONAL METHODS
Thus the quadrupolar contribution decays slowly as 1/R. Higher-order multipole contributions decay as higher powers of 1/R.
First, I calculate the electrostatic potential through construction of the charge density and integrations of Possion’s Equation, the method described in section 2.6.1 and equation 2.29.
Then I use the method of section 2.6.3 and equation 2.39, along with a new plotting method which avoids complications from the Bethe potential by finding the multipolar contributions to the net potential directly. Comparison of the two results demonstrates the simplicity of the new method. My results establish an approximate radial distance at which interfacial contributions die off (charge density averages to zero) providing benchmarks for other studies to
find accurate zeroing of the electrostatic potential in water. Additionally, I find the importance of the quadrupole component due to the local surface around the cavity and the cancelling dipole contributions between the two interfaces. This finding provides an explanation for the wide differences in interfacial potentials for the various water models.
3.2 Computational Methods
The TINKER 7.1.2 code152 was used for all the classical molecular dynamics (MD) simulations.
We study three droplet systems with 250, 512, and 1500 SPC/E water molecules that have radii of approximately 12.1 Å, 15.4 Å, and 22.1 Å, respectively. Finite droplet boundary conditions were applied by implementing a half-harmonic bounding potential on only those water oxygens that traversed a distance of 5 Å outside the approximate radius. All simulations employed a 2 fs timestep and the NVT ensemble using the Berendsen thermostat with 0.1 temperature coupling constant. A temperature of T = 260 K was chosen to reduce the number of evaporating waters, while maintaining adequate diffusion. A 60 ps equilibration was performed on the droplet before cavity formation.
73 3.3. RESULTS AND DISCUSSION
To create the cavity, a neutral point particle was placed at the center of the droplet at x = 0 and restrained to the system’s center of mass. Following the method of Shi and Beck147 the
WCA potential153 was used between the neutral particle and water oxygens to push waters out to several different final cavity radii. The potential is given by M(λ):
c /r12 − c /r6 + c /R12 , r ≤ R 12 6 12 c,λ c,λ M(λ) = (3.3) 0 r > Rc,λ E 2c c = λ , c = 12 12 −6 −6 2 6 6 (λ − Rc,λ) Rc,λ
kB T where λ is the final cavity growth radius, Eλ = 2 , and Rc,λ = 1.05λ (the radius at which the
WCA potential is zero). The cavity was created during 10 consecutive 200 ps growth periods
20 using the functional form fλ(γ) = γ M(λ) with γ = 0.1, 0.2, ..., 1.0. After this 2 ns growth period, the droplet with a cavity was equilibrated for 800 ps before the final 3 ns production run.
3.3 Results and Discussion
To investigate the electrostatic origins of the net potential, we prepared a variety of systems with differing cavity and droplet sizes. This allows us to gauge the size-dependent behavior of the local potential from nearby waters and the surface potential from distant waters97 – termed in other literature as the potential due to the solute-liquid interface and the liquid- vapor interface, respectively.143 Presented here are the results of simulations of a 250 SPC/E water system with cavity radii λ = 2.0, 2.5, 4.0, 4.3, 6.0, 6.6 Å, a 512 SPC/E water system with cavity radii λ = 2.0, 2.5, 4.0, 4.3, 6.0, 6.6, 8.0, 10.0 Å, and a 1500 SPC/E water system with cavity radii λ = 2.0, 2.5, 4.0, 4.3, 6.0, 6.6, 8.0, 10.0 Å. These cavity radii range from small
74 3.3. RESULTS AND DISCUSSION length scales similar to many real ions to nanoscopic sized cavities displaying quasi-macroscopic behavior. Previous work has shown that, for radii larger than about 4 Å, the net potential is relatively stable with increasing size (up to about 16 Å radius).154 For cavities below about 4 Å in radius, the net potential shows a strong size dependence due to local water rearrangements that maximize hydrogen bonding around the small cavities.154
Figure 3.1 shows the electrostatic potential profile obtained using average charge density profiles and eq (2.29) for the droplets of 250, 512, and 1500 waters with cavity radii 4.0 Å and a 1500 water droplet with cavity radii of 2.5 and 6.0 Å centered at x=0. From right to left, the first potential shift of -0.600 V is the surface potential φsp (N = 1500, λ = 6.0 Å case, with nearly equal values for the λ = 2.5 and 4.0 Å cases). The second potential shift (from the bulk region into the cavity) is the local potential φlp with a value of 0.275 V (for λ = 6 Å), and the resulting net potential φnp is -0.325 V. The relatively flat region between the two interfaces is extended in the 1500 water system, but small in the 250 and 512 water systems. This implies that relatively large systems are necessary to obtain adequate bulk-like behavior and interface separation. There is clear charge asymmetry near the cavity, reflected by the large variation in potential. This asymmetry presumably exists locally near the liquid-vapor interface, but is smoothed out due to capillary waves.133
The quantity of interest here is the net potential at the center of the cavity, and perhaps higher-order contributions to the free energy due to asymmetry in the potential distribution.147
One method we can use to determine this potential and its multipole components is to calculate the multipole moment densities and use eqs (2.28)- (2.36) to obtain potential profiles (As in
Fig 3.1). As discussed above, there is added complexity when potentials are measured within the region of charge due to the quadrupole trace term.98 Alternatively, we can avoid these issues by calculating the potential outside of the distribution at the cavity center, x = 0, using
75 3.3. RESULTS AND DISCUSSION
0 0