CHAPTER 6
Semiclassical Bohmian Dynamics
Sophya Garashchuk,a Vitaly Rassolov,a and Oleg Prezhdob a Department of Chemistry and Biochemistry, University of South Carolina, Columbia, South Carolina bDepartment of Chemistry, University of Rochester, Rochester, New York
INTRODUCTION
Quantum mechanics lies at the heart of chemistry. It is impossible to under- stand the structure of the Periodic Table, chemical bonding patterns, free ener- gies of chemical reactions, reaction rates and branching ratios, and other chemi- cal phenomena without a quantum-mechanical (QM) description. In particular, the dynamics of molecular systems often involve QM effects such as zero-point energy, tunneling, and nonadiabatic transitions. QM effects are essential for accurate description and understanding of reactions in complex chemical envi- ronments. For example, the zero-point energy stored in the vibrational modes of chemical reactants, products, and transition state species modifies reaction energy barriers. Reaction rates and branching ratios can be affected greatly by such changes. QM tunneling can be critical in proton transfer reactions. However, the conventional methods of solving the time-dependent Schrodinger¨ equation1 scale exponentially with the system size. In addition, the relevant dy- namics occur on a long time-scale.2 Therefore, it is extremely difficult to model QM tunneling in condensed phase chemical systems.3–6 Nonadiabatic dynam- ics involving transitions between different electronic or vibrational energy levels
Reviews in Computational Chemistry, Volume 27 edited by Kenny B. Lipkowitz Copyright © 2011 John Wiley & Sons, Inc.
287 288 Semiclassical Bohmian Dynamics is ubiquitous in photochemistry.7–11 In this process, a chemical system is excited by light, undergoes a nonequilibrium evolution, and ultimately relaxes to the ground state. QM tunneling can be viewed as a particular kind of nonadiabatic transition.12 Several multidimensional quantum approaches have been proposed in- cluding those using basis set contractions,13–15 coherent state representations, 16,17 and mixed quantum-classical strategies.18–24 A trajectory representation of large molecular systems carries a special appeal because of several favor- able factors: (1) the initial conditions of a trajectory simulation can be sampled with Monte Carlo techniques. This allows the exponential scaling of the exact wave function or density matrix with system size to be circumvented. (2) Var- ious degrees of freedom involving light and heavy particles can be treated on equal footing, and quantum-classical separation issues can be avoided.25 (3) Wave functions are highly oscillatory close to the classical → 0 limit. As a result, a trajectory description of heavy particles, such as nuclei, is often more appropriate than a grid or a basis set representation. (4) Classical equations of motion are simple to solve. Numerous molecular dynamics techniques26 for propagating classical trajectories are applied routinely to chemical systems composing hundreds of thousands of atoms. Incorporating the dominant QM effects caused by wave function localization27,28 constitutes a challenge for the trajectory methods. Such effects can be considered naturally by representing wave functions in terms of ensembles of trajectories. In comparison, semiclas- sical methods commonly use independent trajectories.29–31 A great interest in the Bohmian interpretation of quantum dynamics has been witnessed during the last decade. In particular, its potential to generate computational tools for solving the time-dependent Schrodinger¨ equation has attracted considerable attention. The Bohmian formulation of quantum dynam- ics promises a better than exponential scaling of the computational effort with system dimensionality. It also offers a convenient approach for mixed quantum- classical descriptions of large chemical systems. The Madelung-de Broglie– Bohm formulation of the time-dependent Schrodinger¨ equation has a long his- tory dating back to the birth of quantum mechanics.32,33 It gained a wider recognition after David Bohm34 used it to develop an alternative interpretation of quantum mechanics. This review describes the semiclassical methodologies inspired by the Bohmian formulation of quantum mechanics. These methods are designed to represent the complex dynamics of chemical systems. The review is constructed as follows: the next section introduces the Madelung-de Broglie–Bohm formalism. This is done by drawing an analogy with classical mechanics and explicitly highlighting the non-classical features of the Bohmian mechanics. The nonclassical contributions to the momentum, energy, and force are introduced. The fundamental properties of Bohmian quan- tum mechanics—the conservation and normalization of the QM probability, the computation of the QM expectation values, properties of stationary states, and behavior at nodes—are discussed. Several ways to obtain the classical limit The Formalism and Its Features 289 within the Bohmian formalism are considered. Mixed quantum-classical dy- namics based on the Bohmian formalism is derived and illustrated with an example involving a light and a heavy particle. At this point, the Bohmian rep- resentation is used as a tool to couple the quantum and classical subsystems. The quantum subsystem can be evolved by either Bohmian or traditional techniques. The Ehrenfest approach is the most straightforward and common quantum- classical approach, and it is the starting point for other quantum-classical for- mulations. The Bohmian formulation of the Ehrenfest approach is used to derive an alternative quantum-classical coupling scheme. This resolves the so-called quantum backreaction problem, also known as the trajectory branching prob- lem. Next, the partial hydrodynamic moment approach to coupling classical and quantum systems is outlined. The hydrodynamic moments provide a con- nection between the Bohmian and phase-space descriptions of quantum me- chanics. The “Independent Trajectory Methods” Section describes approaches based on independent Bohmian trajectories. It includes a discussion of the derivative propagation method, the Bohmian trajectory stability approach, and Bohmian trajectories with complex action. Truncation of these hierarchies at the second order reveals a connection to other semiclassical methods. The focus then shifts toward Bohmian dynamics with globally approximated quantum potentials. Separate subsections are devoted to the global energy-conserving approximation for the nonclassical momentum, approximations on subspaces and nonadiabatic dynamics. Each approach is introduced at the formal theo- retical level and then is illustrated by an example. The Section “Towards Reac- tive Dynamics in Condensed Phase” deals with computational issues including numerical stability, error cancellation, dynamics linearization, and long-time behavior. The numerical problems are motivated and illustrated by consid- ering specific quantum phenomena such as zero-point energy and tunneling. The review concludes with a summary of the semiclassical and quantum- classical approaches inspired by the Bohmian formulation of quantum mechan- ics. The three appendices prove the quantum density conservation, introduce quantum trajectories in arbitrary coordinates, and explain optimization of sim- ulation parameters in many dimensions.
THE FORMALISM AND ITS FEATURES
The Trajectory Formulation For simplicity, let us start with a derivation of the Bohmian equations in one spatial dimension x for a particle of mass m. For notation clarity, let us use ∇ to denote differentiation with respect to x. Arguments of functions will be omitted where unambiguous. Differentiation (or Derivatives) with respect to a variable other than x will be indicated as a subscript; for example, ∇c denotes differentiation with respect to c. The multidimensional generalization to an 290 Semiclassical Bohmian Dynamics arbitrary coordinate system is given in Appendix B. The conventional form of the time-dependent Schrodinger¨ equation is 2 ∂ (x, t) − ∇2 + V (x, t) = ı [1] 2m ∂t
After Madelung, the complex time-dependent wave function is represented in polar form as = ı (x, t) A(x, t) exp S(x, t) [2] where A(x, t) and S(x, t) are real functions. Substitution of Eq. [2] into Eq. [1], division by (x, t), separation into real and imaginary parts, and a few simple manipulations results in a system of two equations
∂S(x, t) 1 =− (∇S(x, t))2 − V − Q [3] ∂t 2m ∂(x, t) 1 =−∇ ∇S(x, t)(x, t) [4] ∂t m
In Eq. [3], the term Q denotes what Bohm called the “quantum mechanical potential,”
2 ∇2A(x, t) Q =− [5] 2m A(x, t)
The quantum potential Q enters the equation on par with the external “classi- cal” potential V = V(x, t), which is generally a function of x and t as well as Q = Q(x, t). In Eq. [4] (x, t) is the wave function density
(x, t) = A2(x, t) [6]
With identification of the probability flux as 1 ∗ j(x, t) = (x, t) ∇S(x, t) = (x, t)∇ (x, t) [7] m m
Equation [4] becomes the usual continuity equation. Analogy with fluid me- chanics suggests the name “hydrodynamic” formulation of the Schrodinger¨ equation. Note that Eqs. [3] and [4] are formally equivalent to the original Schrodinger¨ Eq. [1] except that the polar form Eq. [2] is problematic at the nodes of the wave function. At the nodes, the phase S is undefined, A(x, t) = 0, The Formalism and Its Features 291 and Q is generally singular. The singularity in Q cancels for excited eigenstates as will be explained. Equations [3] and [4] describe the flow of the probability fluid through the stationary points x. Transition to the trajectory framework is made by identifying
p(x, t) =∇S(x, t) [8] and switching to the Lagrangian frame of reference
d ∂ p = + ∇ [9] dt ∂t m
Subsequently, the subscript t will be used to define trajectory-dependent quan- tities, such as xt and pt, for the trajectory position and momentum at time t. The subscript 0 will denote the initial values of these quantities at time t = 0. The action function and density along the trajectory (xt,pt) will be denoted as S(xt) and (xt). Variables without subscripts will refer to functions of coordi- nate x and time t. For example, S(x, t) is the phase of a wave function at time t, whereas S(xt) is the action function computed along the trajectory described by the position xt and momentum pt. Differentiation of Eq. [3] with respect to x gives Newton’s equations of motion for a trajectory characterized by momentum pt and position xt
dx p t = t [10] dt m dpt =−∇(V + Q) = [11] dt x xt
In the Lagrangian frame of reference, Eq. [3] becomes the quantum Hamilton- Jacobi equation
2 dS(xt) pt = − (V + Q) x=x [12] dt 2m t
As easily seen, Eqs. [10]–[12] are the standard equations of classical mechanics. These equations fully define the evolution of the wave function once the initial momenta of the quantum trajectories are defined according to Eq. [8]. The quantities xt and pt fully define the quantum trajectory. Quantum effects are incorporated into its behavior through the nonlocal quantum force. The force depends on the wave function amplitude and its derivatives up to the 292 Semiclassical Bohmian Dynamics third order. Often it is useful to consider an additional function attributable to the quantum trajectory, namely, the nonclassical momentum component r(x, t),
∇A(x, t) r(x, t) = [13] A(x, t)
Formally, it is complementary to the classical component p(x, t) because both result from the action of the QM momentum operator on the wave function given in the polar form [2] ∇A(x, t) pˆ = −ı +∇S(x, t) = (−ır(x, t) + p(x, t)) [14] A(x, t)
The quantum potential expressed in terms of r is
2 Q =− r2(x, t) +∇r(x, t) [15] 2m
The average value of Q can be termed the “quantum energy,” which using differentiation by parts in Eq. [15], is equal to
2r2 Q= [16] 2m
In particular, Q is one half of the zero-point energy for the ground state of the harmonic oscillator. The time-dependence of r can be derived from Eq. [4]. Combined with Eqs. [11] and [5], it gives the following evolution equations, which emphasize the common structure of the differential operator on the right-hand-side (RHS): ∇ dpt 2 m∇V x=x + m = rt + ∇rt [17] t dt 2 drt ∇ −m = rt + ∇pt [18] dt 2
Features of the Bohmian Formulation Conservation of Probability and Normalization Transformation of Eq. [4] into the Lagrangian frame of reference gives the evolution of the wave function density d ∂ p (x ) = + t ∇ (x ) =−∇p × (x ) [19] dt t ∂t m t t t The Formalism and Its Features 293
It follows from Eq. [19] that, in closed systems, the probability of finding a particle in the volume element dxt associated with each quantum trajectory, the trajectory “weight,” remains constant in time,35
dw(x ) w(x ) = (x )dx , t = 0 [20] t t t dt
This is consistent with the standard “classical” continuity equation. The cor- responding multidimensional derivation is given in Appendix A. Therefore, all QM effects in the evolution of quantum trajectories result from the quantum force, Fq =−∇Q, acting on the trajectories in addition to the classical force, Fcl =−∇V. The quantum force is responsible for wave-packet delocalization, tunneling, over-the-barrier reflection, resonances, interference, and zero-point energy in bound systems. Conservation of the trajectory “weight” implies that the Bohmian trajectories define the most efficient grid representation for the wave function with time-dependent grid points. The wave function density will remain negligible at these time-dependent grid points provided that it was negligi- ble at time t = 0. Equation [20] also helps to interpret the wave function using the time-dependence of the trajectory positions; because (x, t)is single-valued, the quantum trajectories cannot cross. Wide separation of initially equidistant trajectories indicates regions of low wave function density. Conversely, closely spaced trajectories correspond to high wave function density. A Gaussian wave packet evolving in a harmonic potential provides a sim- ple illustration of quantum trajectory dynamics. The time-dependence of (x, t) is analytic,36 and the trajectories can be constructed easily. The center of the wave packet moves purely classically, whereas the time dependence of the over- all wave packet is influenced by the quantum force, as illustrated in Figure 1. A comparison of the trajectories propagated with and without the quantum potential Q demonstrates the noncrossing rule—classical trajectories cross at focal points, whereas the quantum trajectories do not cross. The noncrossing rule is a manifestation of the single-valued wave function and of the Heisenberg uncertainty principle. An initially wide Gaussian wave packet with the initial width equal to 25% of the coherent value, Figure 1b, produces a small quantum potential Q. In this case, the dynamics of the quantum and classical trajectories are very similar except at the focal points. In contrast, an initially narrow wave packet with the initial width equal to 200% of the coherent value, Figure 1a, produces a large quantum potential Q. As a result, the quantum and classical dynamics differ at all times. The quantum force acts to make the wave packet “flat.” This is the infinite time limit for a Gaussian wave packet evolving in a constant classical poten- tial. For more complicated classical potentials, with ∇V =/ 0, there will be an 294 Semiclassical Bohmian Dynamics
10 (a)
0 Position
-10
0246
10 (b)
0 Position
-10
0246 Time Figure 1 Quantum (dash) and classical (solid line) trajectories in the harmonic potential describing initially (a) narrow and (b) wide Gaussian wave-packets.
intricate interplay between the classical and quantum forces. The interplay gives rise to all QM effects.
The Quantum Trajectory Ensemble, Expectation Values, and Energy To solve the time-dependent Schrodinger¨ equation using trajectories, or sim- ply to visualize the wave function dynamics, an ensemble of trajectories is initialized at t = 0. For each initial position x = x0, the trajectory weight, 2 w = A (x, 0)dx0, and the classical momentum, p0 =∇S(x, 0), are determined from the initial wave function. The trajectories are propagated in time accord- ing to Eq. [10] in combination with either Eq. [11] or Eq. [8]. Equation [11] requires evaluation of A(x, t) and its derivatives through the third order, whereas Eq. [8] needs ∇S(x, t). Once the wave function is represented in terms of trajectories, the expec- tation values of x-dependent operators can be computed readily. This is done The Formalism and Its Features 295 by integrating over the time-dependent trajectory positions. In a discretized (i) trajectory representation, xt , the integration is replaced with the summation = = (i) oˆ (x) t o(x)(x, t)dx o(xt )wi [21] i where the index i enumerates the trajectories. It can be seen easily that the nor- malization of the wave function is conserved; in the Eq. [21], this corresponds to the unit operator oˆ = 1. Operators dependent on p can be evaluated in the same fashion. For example, the kinetic energy is 1 (i) 2 Tˆ t = (p ) w [22] 2m t i i
Note that the total energy of the wave function, or equivalently of the quantum trajectory ensemble, is conserved. However, the energies of individual quantum trajectories generally do not remain constant. The quantum potential is respon- sible for the energy redistribution within the quantum trajectory ensemble.
Stationary States and Behavior at the Nodes A stationary state is a special solution of the time-dependent Schrodinger¨ equa- tion. It can be written as a product of the spatial and temporal factors. Substi- tuting the polar form of the wave function, = − ı (x, t) exp S(t) A(x) [23] into the time-dependent Schrodinger¨ equation, the division of Eq. [23] by , and the separation of variables gives the usual time-independent Schrodinger¨ equation
dS(t) = E [24] dt
2 ∇2A(x) − + V = Q + V = E [25] 2m A(x) where E is the energy eigenvalue. The solution for the phase is S(t) = S(0) + Et. In the quantum trajectory language, an eigenfunction is a wave function with a particular condition on its amplitude. The amplitude of an eigenfunction gives the quantum potential Q which differs from the negative of classical potential −V only by a constant. Therefore, the quantum force corresponding to an eigenfunction exactly cancels the classical force. Given zero initial momenta, (i.e., ∇S = 0) the trajectory positions do not change with time. 296 Semiclassical Bohmian Dynamics
This picture of stationary trajectories also applies to excited eigenstates implying that at the nodes where A(x) = 0 the singularity in the quantum po- tential, given by Eq. [5], always cancels. For example, for the eigenstates of the harmonic oscillator, V = mω2x2/2, the ground-state wave function, mω (x) = exp − x2 [26] 0 2 substituted into Eq. [25] gives E0 = ω/2. The first excited state wave function 1(x) = x0(x) gives E1 = 3ω/2 and so on. Note that the wave function normalization factors have been omitted for clarity. In contrast, for nonstationary wave functions with nodes, the singularities in Q do not cancel. In general, this makes the direct numerical solution of Eqs. [4] and [12] impractical. Nevertheless, the behavior of the quantum trajectories is very intuitive; the trajectories “flow” coherently, avoid the nodes, and never cross. The trajectories corresponding to the time-evolution of a linear combination of 0 and 1 are shown on Figure 2 around the density node at t = 2.5. Avoiding the nodal region, Figure 2a, is accomplished by rapid changes in the trajectory momenta, Figure 2b. As a consequence, numer- ical implementation of such unstable dynamics is very expensive. Many examples of the quantum trajectory dynamics can be found in the book by Holland.37
(a) (b)
5 1
0 0 Position Momentum
-1
-5
12341234 Time Time Figure 2 Bohmian dynamics in the presence of the density nodes: (a) position of the trajectories as a function of time; (b) the corresponding momenta for selected trajectories. Position and momentum of a trajectory are shown with the same line styles on both panels. The Formalism and Its Features 297
The Classical Limit of the Schrodinger¨ Equation and the Semiclassical Regime of Bohmian Trajectories The Bohmian form of the time-dependent Schrodinger¨ equation gives a straightforward route to classical mechanics. In the heavy particle limit, m → ∞, or equivalently, when a typical action becomes large compared with Planck’s constant, → 0, the quantum potential Q vanishes. This is consistent with the representation of a particle in terms of a localized wave function (i.e., a wave- packet). In the classical limit, the center of the wave packet moves along a classical trajectory, and the changes in the wave-packet width can be neglected. An alternative connection between Bohmian, classical, and semiclassical dynamics is through the Wentzel–Kramers–Brillouin (WKB)38 treatments that are based on the expansions of the exponentiated wave function, as in Eq. [60], which is described later. The traditional semiclassical condition is based on the WKB approximation to solutions of the time-independent Schrodinger¨ equa- tion. This condition states that the action function must be much larger than Plank’s constant. The classical momentum p entering the action function is defined as pWKB = 2m(E − V), pdx [27]
Identification of the classical and nonclassical momenta, Eqs. [8] and [13] re- spectively, suggests a similar criterion that is applicable to time-dependent wave functions: |r| |p| [28] or in terms of the energy given in Eq. [15],
p2 |Q| [29] 2m
The momentum condition of Eq. [28] is more convenient than the energy condition Eq. [29] the former is expressed in terms of simple quantities, which are linear in the semiclassical picture of a moving particle (i.e., for Gaussian wave-packets). In the context of trajectory dynamics, the momentum semiclassical condi- tion Eq. [28] is more general than the WKB expression Eq. [27]. This is because it is not based on a particular approximate solution to the Schrodinger¨ equation. Moreover, the condition of Eq. [28] is more convenient because it is expressed in terms of r and p, which are natural attributes of quantum trajectories. For semiclassical systems with small Q, the Bohmian momentum and the WKB momentum are close to each other. Therefore, the momentum semiclassical condition and the WKB expression are related closely. 298 Semiclassical Bohmian Dynamics
According to Eq. [28], the semiclassical approximation breaks down near wave function nodes where A(x, t) = 0. This so-called “node problem” leads to singular forces acting on quantum trajectories and causing numerical insta- bilities. A similar breakdown of the WKB approximation occurs near classical turning points. In the context of purely classical trajectories, this problem was dealt with by developing uniform semiclassical methods.30,39 A general-purpose semiclassical method based on quantum trajectories also must satisfy Eq. [28] in a uniform sense (i.e., for all points in the coordinate space). This require- ment motivates the development of approximate quantum potentials outlined in the section, “Global Energy Conserving Approximation of the Nonclassical Momentom”. Approximate quantum potentials are defined through the lin- earization of the nonclassical momentum via averaging over the wave function density. The relevant semiclassical condition becomes
|r| |p| [30]
The singularities in r have negligible contributions to the dynamics because of vanishing wave function density. A general semiclassical method must satisfy the semiclassical condition at all times during time evolution; therefore, it cannot be defined for any specific form of the initial wave function. In practice, this means that approximations must be made only for quantities that are negligible in the m →∞or → 0 limits. In particular, removing singularities in the quantum potential by con- straining density, phase, or momentum generally would violate the semiclassi- cal condition. The approximate quantum potential method of “Global Energy- Conserving Approximation of the Nonclassical Momentum” constrains the functional form of the nonclassical momentum r(x, t), which enters Eq. [28] with the prefactor. At the same time, the density itself remains unconstrained. The traditional semiclassical methods, such as WKB and the Van Vleck– Gutzwiller propagator,38,40,41 as well as the independent Bohmian trajectory methods, are defined through the expansion of the solution to the Schrodinger¨ equation. Independent Bohmian trajectory methods, such as the derivative propagation method, the Bohmian trajectory stability method, and Bohmian mechanics with complex action, are discussed in the section on “The Inde- pendent Trajectory Methods.” The -expansion converges to the exact result. Methods that are not based on analytic solutions, such as the approximate quantum potential method, can be considered semiclassical if they can be im- proved systematically in the limit of large mass for an arbitrary physically rea- sonable initial wave function and kinetic energy density. The kinetic energy den- sity is defined as −2∇2 /(2m ). “Improved systematically” implies that there is a general, unambiguous numerical prescription for convergence toward the exact solution. The approximate quantum potential approach can be improved systematically if the linearization of the nonclassical momentum is accom- plished over subspaces,42 or if r(x, t) is represented in terms of a complete basis. The Formalism and Its Features 299
Using Quantum Trajectories in Dynamics of Chemical Systems Conceptually, the quantum trajectory formalism has been extended to nonadiabatic dynamics,43–45 the phase-space representation, and the density matrix approaches.46–53 There exist developments on imaginary time propa- gation,54 complex space Bohmian trajectories,55 and dynamics that are based on the bipolar rather than polar decomposition of wave functions.56,57 On the practical side, quantum trajectories have been used in the- oretical and computational chemistry for three distinct purposes: (1) to interpret the wave function computed by the conventional wave function propagation techniques; (2) to monitor the wave function density flow for dynamical grid adjustments; and (3) to solve the time-dependent Schrodinger¨ equation, and to obtain (x, t) or quantities of interest, directly from the trajectories. Quantum trajectories have been used to interpret and to draw quantum- classical analogies in the area of surface scattering. Phenomena including Fresnel and Fraunhofer regimes, rainbow scattering, the quantum Talbot effect, and others have been explored.58–61 Some ideas from the Bohmian dynamics lead to the moving grid techniques62 in which positions of grid points are time-dependent but are different from the Bohmian trajectories. The numerical goals of moving grid techniques are to gain stability of the dynamics of grid points and to improve accuracy in derivative evaluations. Instead of moving the grid points, grids can be optimized by adding or eliminating the grid points by reconstructing Bohmian trajectories at grid edges.63 A similar criterion is used in the ab initio wave-packet dynamics of Iyengar to optimize the wave-packet representation and to minimize ab initio evaluations of classical forces.64–66 Using quantum trajectories as a practical way of solving the multidimen- sional time-dependent Schrodinger¨ equation is an exciting prospect. Several high-dimensional applications of the exact quantum trajectory method, includ- ing up to 200 degrees of freedom, have been reported.67,68 There, the quantum force is evaluated on the fly with the moving least-squares fitting of the wave function amplitude. However, a general implementation of the exact numerical Bohmian trajectory technique is difficult, even for low-dimensional systems. Complications develop as a result of the singularities in the quantum potential. 62,69–71 These problems motivated the “independent trajectory” implementa- tions based on the Taylor expansion of the equations of motion truncated at a low order. Both real-valued54,72 and complex-valued trajectories55,73 have been used. The approximate quantum potential (AQP) approach35,74,75 in- volves propagation of the trajectory ensemble together with a global evalua- tion of the quantum force from the moments of the trajectory distribution. This “mean-field” type of approximation gives quantum force for all trajectories si- multaneously. 300 Semiclassical Bohmian Dynamics
In the remainder of this review, the focus will be on the semiclassical and approximate implementations of Bohmian mechanics that have the greatest potential for high-dimensional chemical applications.
BOHMIAN QUANTUM-CLASSICAL DYNAMICS
Dynamics of most chemical reactions are typically very complex for a fully quantum-mechanical analysis. Fortunately often, it is possible to distin- guish between particles, such as electrons and protons, that require a quantum description and particles, such as heavy nuclei, which can be described accu- rately using classical mechanics. When quantum particles remain in the same quantum state throughout the reaction, the Born–Oppenheimer (adiabatic) ap- proximation is invoked. The quantum state merely provides an external poten- tial for the classical dynamics leading to adiabatic molecular dynamics.8,76–79 The Born-Oppenheimer approximation is valid, for example, for thermally acti- vated nuclear rearrangements proceeding in the ground electronic state. Many other types of chemical reactions involve several quantum states. Examples include photochemical reactions,80,81 transfers of electrons,11,82 protons,83,84 spins,85 energy86–88 and quantum phase,89 electron-vibrational relaxation pro- cesses,9,10,90,91 and solvation dynamics.92,93 These phenomena extend beyond the Born–Oppenheimer approximation and are modeled by the nonadiabatic generalizations of molecular dynamics.94–104 Coupling between quantum and classical degrees of freedom constitutes the key question in mixed quantum-classical approaches. It raises central issues that do not admit unique solutions. Numerous coupling schemes have been proposed ranging from formal mathematical solutions20,105–110 to specific al- gorithms that have been applied to many problems in chemistry, physics, and biology.9–11,80–84,87–92 Historically, the first and the most straightforward of the quantum-classical approaches is based on the Ehrenfest theorem. The theo- rem states that the equations of motions for the average values of the quantum position and momentum operators coincide with the classical equations of mo- tion.99,111 This leads to the mean-field approximation in which the classical variables are coupled to the expectation values of the quantum observables. If the quantum system remains in a single state, then the Ehrenfest approach re- duces to adiabatic molecular dynamics. In general, the quantum system forms a superposition of several states, and the classical dynamics evolve in the mean- field quantum potential. The average Ehrenfest trajectory is inappropriate when several reaction channels exist and involve substantially different potential en- ergy surfaces.94,96 In a corresponding quantum description, the wave packets split and follow different reaction channels. The branching difficulty in the coupling of quantum and classical mechanics is known as the quantum back- reaction problem. Most often, it is resolved by surface hopping94–97,100,101 in Bohmian Quantum-Classical Dynamics 301 which classical trajectories are designed to branch according to a specific al- gorithm. Other, more computationally demanding, quantum-classical approx- imations dealing with the trajectory branching include the multiconfiguration mean-field theory,20,98,103 partial Wigner transform dynamics,20,108 and semi- classical treatments.104 The Bohmian interpretation of quantum mechanics34 provides an alternative means of achieving the branching of the classical tra- jectories.24,25 By correlating each classical trajectory with an individual parti- cle, an ensemble of trajectories can be generated. Trajectories associated with different quantum states are represented by different Bohmian particles. Tra- jectories evolve independently and branch as in the fully quantum-mechanical description. Subsequently, we describe the Ehrenfest and Bohmian quantum- classical approaches. Their properties are illustrated by a model representing scattering of a light particle off a surface containing slow phonon modes. The photoinduced electron transfer from a molecule to a semiconductor surface in dye-sensitized semiconductor solar cells11,82 is an example of such a process.
Mean-Field Ehrenfest Quantum-Classical Dynamics Consider a mixed quantum (x) classical (X) system. The quantum Hamil- tonian H(x; X) depends parametrically on the positions of classical particles
2 H(x; X) =− ∇2 + V(x; X) [31] 2m x
The classical subsystem generates an external field contributing to the potential V(x; X) that governs the motion of the quantum subsystem. The total quantum- classical energy is the sum of the quantum-mechanical expectation value of the Hamiltonian [31] with the purely classical kinetic and potential W(X) energies 2 − ∗ MX˙ Eq cl = Eq + Ecl = d3x (x)H(x; X)(x) + + W(X) [32] 2
The wave function (x) evolves according to the time-dependent Schrodinger¨ equation ∂(x) 2 i = − ∇ 2 + V(x; X(t)) (x) [33] ∂t 2m x in which the potential V depends on time through the dynamics of classical variables X(t). The evolution of the classical coordinates obeys the Newton equation
q MX¨ =−∇XW(X) + F [34] 302 Semiclassical Bohmian Dynamics which contains the quantum force Fq in addition to the ordinary classical force −∇XW(X). The definition of the quantum force constitutes the quantum back- reaction problem.20,105–110 The quantum force of the Ehrenfest approach is given by the quantum mechanical expectation value of the gradient of the quan- tum Hamiltonian q 3 ∗ F =− d x (x) [∇XH(x; X)] (x) [35]
The Ehrenfest force conserves the total quantum-classical energy in Eq. [32], as established by the time-dependent Hellmann–Feynman theorem.112 The qualitative features of the quantum-classical Ehrenfest approxima- tion are illustrated in Figure 3. For a given wave function of the quantum subsystem, the Ehrenfest force defines a unique classical trajectory, Figure 3b. This feature of the Ehrenfest method is both its major advantage and its dis- advantage. Consider the case in which the quantum-mechanical wave packet, corresponding to the classical subsystem in the Ehrenfest approach, remains localized throughout the time of an experiment. Here, the Ehrenfest force of Eq. [35] generates an optimal, classical description of the wave packet. On the other hand, if the wave packet branches, as illustrated in Figure 3a, then the Ehrenfest approach fails to capture the branching. It asymptotically cannot de- scribe the distinct reaction channels of the classical subsystem associated with different quantum states.
Quantum-Classical Coupling via Bohmian Particles The branching of the classical subsystem is reproduced by the Bohmian quantum-classical approach. This is done by generating an ensemble of classi- cal trajectories correlated with different members of the Bohmian ensemble of quantum particles. Consider the Bohmian formulation of the Ehrenfest force of Eq. [35]. The polar form of the wave function (x) = R(x) exp(ıS(x)/) leads to the following expression for the quantum mechanical expectation value of the Hamiltonian Eq. [31]: 2 ∗ d3x (x) − ∇2 + V(x; X) (x) = d3xR2(x) 2m x (∇ S(x))2 × x + Q(x) + V(x; X) [36] 2m where Q(x) is the quantum potential
2 ∇ 2R(x) Q(x) =− x [37] 2m R(x) Bohmian Quantum-Classical Dynamics 303
Figure 3 Schematic representation of the evolution of a heavy particle moving in the po- tential created by light particles. For example, the X coordinate can represent the bond length of a diatomic molecule. The black solid lines depict the potential energy sur- faces for the molecular ground and excited electronic states. (a) A quantum-mechanical wave-packet describing the heavy particle and shown by the grey dashed line is pro- moted from the ground electronic state to the excited state, as indicated by the grey solid line directly above the dashed line. While in the excited state, the wave-packet moves to infinite X, and the diatomic dissociates. The nonadiabatic coupling between the electronic states causes transfer of a fraction of the wave-packet back to the ground state potential energy surface. This part of the wave-packet returns to the initial state, representing a bound diatomic. Thus, a quantum-mechanical particle branches into sev- eral components, corresponding to different outcomes of the excitation dynamics. (b) In the Ehrenfest approximation, the classical particle cannot split and evolves on a sin- gle trajectory. The mean-field potential energy surface (grey line) is an average of the ground and excited state potentials. The particle evolution corresponds to neither of the quantum-mechanical outcomes; the diatomic can be artificially trapped between the bound and dissociated states. (c) In the Bohmian version of quantum-classical dynamics, the quantum subsystem, e.g. the electrons in the diatomic, is represented by an ensemble of classical-like particles. Each Bohmian particle is coupled independently to the heavy particle. This coupling generates an ensemble of classical trajectories evolving on differ- ent potential energy surfaces. This treatment mimics the quantum-mechanical branching of the wave-packet describing the heavy particle.
The quantum probability distribution is R2(x) = ∗(x)(x). As a result, in Bohmian mechanics, the quantum energy Eq. [36] is interpreted as the energy of an ensemble of particles with the probability distribution R2(x). The energy of 2 each particle is equal to [(∇xS(x)) /2m + Q(x) + V(x; X)]. The Ehrenfest force, 304 Semiclassical Bohmian Dynamics generated by the quantum subsystem and acting on the classical subsystem, takes the following form in the Bohmian representation: q 3 2 F =− d xR (x)∇XV(x; X) [38]
It can be viewed as the average of the forces −∇XV(x; X) resulting from the ensemble of Bohmian particles with the probability distribution R2(x). The ensemble averaged Bohmian force Eq. [38] is identical to the Ehrenfest force of Eq. [35], as indicated by the time-dependent Hellmann-Feynman theorem.112 The theorem states that the time derivative of the expectation value of the quan- tum energy involves only the derivative of the quantum Hamiltonian, provided that the wave function evolves according to the time-dependent Schrodinger¨ Eq. [33]. The Bohmian quantum-classical approach solves the branching problem by moving the d3xR2(x) ensemble averaging outside the quantum-classical dynamics. The initial conditions for the Bohmian ensemble of quantum particles are sampled from R2(x). The averaging is performed only at the final time. In particular, the quantum force acting on the classical subsystem at any given instance does not involve the integration over R2(x) as in Eq. [38]. The quantum force is calculated for a single representative of the Bohmian ensemble. This treatment of the quantum-classical coupling generates a distribution of classical trajectories correlated with different Bohmian particles, Figure 3c. As a result, the classical trajectories evolve differently depending on whether the correlated Bohmian particles correspond to the excited state wave function, Figure 3c, or to the ground state. The Bohmian quantum-classical simulation runs according to the follow- ing algorithm. First, initial conditions for the wave function and classical tra- jectories are chosen in the usual manner. Positions x of Bohmian particles are sampled from the initial distribution R2(x). For each initial coordinate X of the classical particle, an ensemble of initial coordinates x for the Bohmian particles is sampled from R2(x). Each Bohmian particle is correlated with a separate copy of the classical subsystem. Second, for each member of the quantum-classical ensemble, the wave function is propagated by the time-dependent Schrodinger¨ Eq. [33]. Simultaneously, the classical trajectory is evolved by the Newton Eq. [34] with the quantum force
q F =−∇XV(x; X) [39]
Note that the quantum force depends on the position of the Bohmian particle x. The trajectory of the Bohmian particle is propagated either using the Newton equation, including the quantum potential Eq. [37]
mx¨ =−∇x[Q(x) + V(x; X)] [40] Bohmian Quantum-Classical Dynamics 305 or, equivalently, directly by
x˙ =∇xS/m [41]
The first option can be used with the semiclassical schemes for propagating the Bohmian ensemble described in the sections “The Independent Trajectory Methods” and “Dynamics with the Globally Approximated Quantom Poten- tial” of this review. The second option is advantageous if the time-dependent wave function is available by a direct quantum-mechanical propagation. Third, the results are averaged over the ensemble of the Bohmian quantum-classical trajectories. The Bohmian quantum-classical method is defined by Eqs. [33], [34], [39], and [40]. The fully classical limit for both subsystems is achieved easily by setting → 0 in Eq. [37] and eliminating the quantum potential from Eq. [40]. An alternative derivation of this approach is given in Refs. 24 and 113. The derivation starts with a fully quantum description of both subsys- tems. The quantum potential then is dropped from the equation of motion for the classical subsystem.
Numerical Illustration of the Bohmian Quantum-Classical Dynamics The Bohmian quantum-classical approach is illustrated here with a model intended originally as a simplified representation of gaseous oxygen interacting with a platinum surface.114,115 Alternatively, it can be viewed as a model for the photoinduced electron transfer in a molecular chromophore adsorbed on a surface of a solid-state bulk material. Systems of this type form the basis for dye-sensitized semiconductor solar cells, also known as Gratzel¨ cells,11,82 and various molecular electronics devices.116 The model consists of a light particle x with mass m colliding with a heavier particle X with mass M. The heavy particle is bound to an immobile surface, Figure 4. In the molecule-bulk electron transfer process, the light particle can be viewed as the electron coming from the molecule and scattering off the bulk surface containing a phonon mode. The total Hamiltonian for the system is given by
H(x; X) = T1(x) + V1(x) + T2(X) + V2(X) + V(x, X), with [42a] −2b(x−c) −b(x−c) V1(x) = a[e − 2e ] [42b] 1 V (X) = M2X2 [42c] 2 2 − − V(x, X) = Ae B(x X) [42d] where T1 and T2 are the kinetic energy operators. The harmonic potential V2 describes the interaction of the heavy particle with the surface. The Morse 306 Semiclassical Bohmian Dynamics
Figure 4 A model illustrating the advantages of the Bohmian quantum-classical dy- namics over the Ehrenfest approach. The top panel depicts the system. The light particle approaches and scatters off a surface which contains a phonon mode involving the heavy particle. The Hamiltonian and its parameters are given in Eq. [42a] and Table I, respec- tively. The bottom panels show the time-dependent probability for the light particle to move a certain distance away from the surface following the scattering event. The data shown in the two panels differ in the initial kinetic energy of the light particle E. Quan- tum mechanically, the light particle has a 100% probability to leave the surface after a sufficiently long time. The light particle wave-packet splits, and a part of it remains tem- porarily trapped with the heavy particle. In the exact solution, the trapped part of the wave-packet follows the scattered part and eventually decays. The Ehrenfest approach errs both in the scattering onset and the asymptotic scattering probability, as highlighted by the boxes in the middle panel. The classical treatment of the phonon mistreats zero- point energy and allows transfer of the phonon energy to the light particle, accelerating the scattering. At longer times, the lack of branching creates an artificial constraint on the energy exchange between the light and heavy particles. The excess energy transferred from the heavy to the light particle during the early evolution leaves the heavy particle with insufficient energy to continue promoting the scattering of the light particle. Bohmian Quantum-Classical Dynamics 307
Table 1 Parameters Used in Simulation of the Scattering Problem, Eq. [42a] and Figure 4 m 1 amu a 700 kJ/mol M 10 amu b 5.0 A˚ −1 4 × 1014 s−1 c 0.7 4 A10kJ/mol x0 6.0 A˚ B 4.25 A˚ −1 0.5 A˚
potential V1 describes the interaction of the light particle with the surface. The two particles interact by the exponentially repulsive potential V. Parameters particular to the simulation are provided in Table 1 and are the same as in Refs. 25, 115 and 117. Initially, the light particle is moving toward the heavy particle. The light particle is described by a Gaussian wave-packet