betaFeedback

WorksPhysics and Astronomy

Surface Hopping Dynamics on Adiabatic States Propagated with Real-Time TDDFT

Abstract

Real-time time-dependent density functional theory (RT-TDDFT) is widely used to simulate ultrafast electron dynamics, but it is usually coupled to the nuclei through the mean-field Ehrenfest approximation, which cannot describe wavepacket branching and violates detailed balance. Surface hopping avoids these problems, but has not previously been combined with RT-TDDFT because real-time propagation does not directly provide the necessary adiabatic states. In this work we introduce an energy-scaled time-dependent Kohn–Sham (TDKS) propagation scheme, motivated by the quantum adiabatic theorem. Scaling widens the separation between electronic and nuclear timescales, so that each propagated state follows the adiabatic surface selected by its initial condition: either the ground state or a non-Aufbau self-consistent-field (ΔΔSCF) solution. The propagation thereby avoids the convergence failures of static SCF solvers in regions of strong nonadiabatic coupling. Because the independently propagated states are complex-valued and not mutually orthogonal, we reformulate the mapping approach to surface hopping (MASH) in a complex adiabatic basis and introduce a dynamical transition orbital (DTO) gauge that restores a well-defined particle–hole representation for evaluating time-derivative couplings and nonadiabatic coupling vectors. Applied to the photoisomerization of trans-diimide, the scheme converges systematically as the scaling factor is increased and yields an excited-state lifetime in good agreement with previous nonadiabatic simulations based on static SCF.@ These results establish real-time propagation as an efficient and numerically stable route to the adiabatic states and couplings required for first-principles surface hopping.

Introduction

Simulating photoinduced chemical reactions requires going beyond the Born–Oppenheimer approximation because photoexcited molecules readily access regions of strong nonadiabatic coupling such as the vicinity of conical intersections. [1] The simplest nonadiabatic dynamics scheme employs the Ehrenfest approximation, which propagates the nuclei on a mean-field average of the potential-energy surfaces. [2] However, its mean-field character prevents nuclear trajectories from branching onto distinct product surfaces [3] and can lead to severe violations of detailed balance. [4] Surface-hopping methods, such as fewest-switches surface hopping (FSSH)[5] and the mapping approach to surface hopping (MASH),[6, 7] instead propagate each trajectory on a single active adiabatic surface and define a rule for transitioning between states. In this way, an ensemble of surface-hopping trajectories can branch onto different surfaces after passing through a nonadiabatic coupling region, which typically leads to more reliable results than Ehrenfest dynamics.

Most surface-hopping methods require the adiabatic electronic states ΦN\Phi_N and the energies ENE_N, forces and nonadiabatic couplings derived from them.[8, 9] These states are formally defined as the eigenstates of the electronic Hamiltonian,

H^elΦN=ENΦN(1)\hat{H}_{\mathrm{el}}\Phi_N = E_N\Phi_N \tag*{(1)}

For model Hamiltonians, the states are simply calculated by matrix diagonalization, whereas obtaining them on the fly from first principles is considerably more demanding. Multireference wavefunction methods such as complete-active-space second-order perturbation theory (CASPT2){sup}10 and multireference configuration interaction (MRCI){sup}11 can describe excited states and conical intersections accurately, but their computational cost restricts the accessible system size. Density functional theory (DFT), [12, 13], with the Kohn–Sham (KS) representation of the density by an auxiliary noninteracting determinant, [14], provides a more affordable starting point for the ground state. The challenge is to obtain the energies, forces and couplings of excited states reliably as the nuclei move through strongly coupled regions.

DFT offers several ways to treat excited states within a static picture, in which calculations are performed independently at each time step. Linear-response time-dependent density functional theory (LR-TDDFT), [15, 16], which supplies a manifold of excitation energies and response states, is widely used with surface hopping. [17] Its cost is, however, higher than that of ground-state KS-DFT, and with commonly used adiabatic exchange–correlation kernels, it does not describe intersections involving the ground state correctly. [18]

The Δ\DeltaSCF method instead approximates an excited state by a separate self-consistent-field (SCF) calculation with a non-Aufbau occupation, [19], at a cost comparable to a ground-state calculation, and is often more accurate than LR-TDDFT for charge-transfer, doubly excited states and conical intersections. [20, 21, 22] Although Δ\DeltaSCF has been combined with surface hopping, [24, 25] [23], variational collapse [26] and switching between competing SCF solutions remain practical difficulties. The most widely used strategy is the maximum overlap method (MOM), [27], which at each SCF iteration occupies the orbitals that overlap most with those of the previous iteration. Other optimizers target individual excited states directly. [20, 28, 29, 30]

Along a trajectory, however, the overlap criterion must be reapplied at every geometry, and near degeneracies, where several orbitals overlap comparably with the reference, it can incorrectly switch to a different solution. Related state-specific approaches include restricted open-shell Kohn–Sham (ROKS), [31, 32], σ\sigma-SCF, [33], and constrained DFT. [34, 35, 36]

In contrast to these static, state-specific approaches, real-time TDDFT (RT-TDDFT) propagates the KS orbitals directly in the time domain, [37, 38, 39], giving access to strongly nonequilibrium processes [40, 41, 42, 43] and to linear and nonlinear optical response. [44] Real-time propagation can be more efficient and more stable than the static SCF approaches discussed above. [45, 46]

The propagated density does not directly supply separate adiabatic-state energies, forces and couplings, however, so coupled electron–nuclear applications have largely relied on Ehrenfest dynamics.[2, 47, 48, 49, 50, 51, 52, 53, 54, 55] Extensions beyond this level differ mainly in what takes the place of the adiabatic basis: natural-orbital branching,[56] density-matrix ensembles with decoherence corrections,[57] and approaches based on exact factorization.[58, 59, 60] Another widely used approach describes transitions between instantaneous eigenstates of the time-dependent Kohn–Sham (TDKS) Hamiltonian,[61, 62] in an independent-electron picture also used for surface hopping at metal surfaces.[63] However, these single-particle states are not the same as the many-electron adiabatic states of Eq. (1). Implementations such as PYXAID,[64] Hefei-NAMD,[65] and GTSH[66] have enabled simulations of exciton and carrier relaxation in condensed-phase systems.[67] For simplicity, many such calculations use the classical-path approximation, in which the nuclei follow a ground-state trajectory irrespective of the active electronic state.[64] However, photochemistry involving large nuclear rearrangement, such as photoisomerization, requires feedback from nuclear forces resulting from the many-electron adiabatic states.

The concept of propagating electronic states to follow adiabatic surfaces has precedents in Car–Parrinello molecular dynamics (CPMD)[68, 69, 70, 71] and other extended-Lagrangian molecular dynamics,[72] originally developed for ground-state Born–Oppenheimer dynamics. In CPMD, a fictitious mass controls the separation between electronic and nuclear timescales required for adiabatic following.[73] Extensions have been used for propagation on excited-state surfaces.[31, 74, 75, 76] For instance, the nonadiabatic CPMD scheme of Doltsinis and Marx[77] propagates a ROKS excited state alongside a restricted Kohn–Sham ground state and evaluates their coupling from the finite-difference overlap expression of Hammes-Schiffer and Tully.[78] However, this approach assumes orthonormal states, whereas independently generated KS states need not be mutually orthogonal. For open-shell states, mixing of equally occupied active spatial orbitals can also change the constructed many-electron wavefunction without changing the density, as discussed in studies of ROKS.[79, 80] A propagation-based surface-hopping scheme therefore needs both reliable following of the selected SCF branches and a consistent wavefunction representation for evaluating their overlaps and couplings.

Here we combine energy-scaled TDKS propagation with a complex-state formulation of MASH. Motivated by the quantum adiabatic theorem, we multiply the electronic Hamiltonian by a factor α>1\alpha> 1 and propagate each state independently. For a prescribed nuclear path, this scaling is equivalent to traversing the same path at 1/α1/\alpha of the original speed, so that each propagated state follows its adiabatic state more closely (§ 2.3). Note that a similar electronic equation of motion underlies the modified Ehrenfest dynamics of Rubio and coworkers, which uses the opposite scaling regime to permit larger time steps in ground-state simulations. [45, 46] In order to recover the nonadiabatic dynamics that we wish to describe, we use MASH with unscaled energies and forces. MASH is a variant of surface hopping that uses a deterministic algorithm for selecting the active surface and shows a number of benefits over the original FSSH approach. [6, 7] Nonetheless, the implementation of MASH is very similar to that of other surface-hopping approaches: they require energies, forces and time-derivative couplings at each time step, as well as nonadiabatic coupling vectors at hops. Since real-time propagation produces complex states, we recast the MASH equations in a complex adiabatic basis. A dynamical transition orbital (DTO) [81] gauge fixes the particle–hole representation used to construct nearly orthogonal many-electron wavefunctions. Their overlaps then provide the required time-derivative couplings and nonadiabatic coupling vectors.

We demonstrate the scheme for the photoisomerization of trans-diimide, testing state following, wavefunction overlaps and nonadiabatic couplings before performing full RT-TDDFT/MASH dynamics. Increasing α\alpha systematically improves state following through regions where static SCF optimization is unreliable, and the ensemble populations and structural observables converge to results in good agreement with earlier surface-hopping simulations based on static Δ\DeltaSCF calculations. [82] In this proof-of-principle study, the main demonstrated advantages are robustness and numerical stability. For larger systems, where static excited-state SCF calculations become increasingly difficult to converge, the approach is also expected to be more efficient than standard static approaches (as observed in other RT-TDDFT studies of large systems), [45, 46] although a systematic cost comparison is beyond the scope of this work.

We describe the electronic propagation in § 2, the coupled nonadiabatic dynamics in § 3, test the method in § 4, and conclude in § 5.

RT-TDDFT for Adiabatic States

Notation. We first fix the conventions used throughout. Matrices are set in upright boldface (e.g., S\mathbf{S}, H\mathbf{H}, C\mathbf{C}) and vectors in italic boldface (e.g., R\mathbf{R}, P\mathbf{P}, d\mathbf{d}). The letter SS is conventionally used both for the atomic-orbital overlap matrix of the LCAO equations of § 2.5 and for the spin (Bloch) vector of the MASH mapping variables of § 3.1. We retain both usages but write the latter with an arrow, S⃗\vec{S}, with Cartesian components SxS_x, SyS_y, SzS_z, so that the two are distinguished at a glance. For the same reason, the orbital-overlap matrices used to evaluate many-electron overlaps in § 3.3 are denoted O\mathbf{O} rather than the more common S\mathbf{S}. Atomic units are used unless stated otherwise.

Adiabatic States from the Variational Principle

Within the framework of Kohn–Sham density functional theory (KS-DFT), [13, 14] the adiabatic ground state for a fixed nuclear configuration R\mathbf{R} is obtained by minimizing the total electronic energy functional EKS[ρ]E_{\mathrm{KS}}[\rho], which is the sum of the noninteracting kinetic energy, the external-potential energy, the classical Hartree energy, and the exchange–correlation energy, [83]

EKS[ρ;R]=Ts[{ψn}]+∫vext(r;R)ρ(r) d3r+12∬ρ(r)ρ(r′)∣r−r′∣ d3r d3r′+Exc[ρ](2)E_{\mathrm{KS}}[\rho;\mathbf{R}] = T_s[\{\psi_n\}] + \int v_{\mathrm{ext}}(\mathbf{r};\mathbf{R})\rho(\mathbf{r})\,\mathrm{d}^3r + \frac{1}{2}\iint\frac{\rho(\mathbf{r})\rho(\mathbf{r}')}{\lvert\mathbf{r}-\mathbf{r}'\rvert}\,\mathrm{d}^3r\,\mathrm{d}^3r' + E_{\mathrm{xc}}[\rho] \tag*{(2)}

Here, the density ρ(r)\rho(\mathbf{r}) is a function of a single-electron coordinate r\mathbf{r} and is constructed from the occupied KS orbitals:

ρ(r)=∑nfn∣ψn(r)∣2(3)\rho(\mathbf{r}) = \sum_n f_n \lvert\psi_n(\mathbf{r}) \rvert^2 \tag*{(3)}

where fnf_n is the occupation number of orbital nn. Orthonormality of the orbitals is imposed during the minimization through Lagrange multipliers εnm\varepsilon_{nm},

L[{ψn};R]=EKS[ρ;R]−∑n,mεnm(⟨ψm∣ψn⟩−δmn)(4)\mathcal{L}[\{\psi_n\}; \mathbf{R}] = E_{\mathrm{KS}}[\rho; \mathbf{R}] - \sum_{n,m} \varepsilon_{nm}(\langle\psi_m \mid\psi_n \rangle- \delta_{mn}) \tag*{(4)}

Variation with respect to the orbitals gives the KS self-consistent-field (SCF) equations

H^KS[ρ]ψn=εnψn(5)\hat{H}_{\mathrm{KS}}[\rho]\psi_n = \varepsilon_n\psi_n \tag*{(5)}

where H^KS[ρ]=−12∇2+v^ext(R)+v^H[ρ]+v^xc[ρ]\hat{H}_{\mathrm{KS}}[\rho] = -\frac{1}{2}\nabla^2 + \hat{v}_{\mathrm{ext}}(\mathbf{R}) + \hat{v}_{\mathrm{H}}[\rho] + \hat{v}_{\mathrm{xc}}[\rho] is the KS Hamiltonian. Because H^KS\hat{H}_{\mathrm{KS}} depends on ρ\rho and hence on the KS orbitals themselves, Eq. (5) is a nonlinear eigenvalue problem that must be solved iteratively. The resulting potential-energy surface on which the nuclei move is V(R)=EKS[ρ;R]+Enuc(R)V(\mathbf{R}) = E_{\mathrm{KS}}[\rho; \mathbf{R}] + E_{\mathrm{nuc}}(\mathbf{R}), where EnucE_{\mathrm{nuc}} is the nuclear repulsion energy.

In practical implementations, a range of numerical techniques is used to accelerate the convergence of the SCF iteration of Eq. (5). Direct inversion in the iterative subspace (DIIS) [84] extrapolates the KS Hamiltonian from previous iterations, whereas orbital-optimization methods parametrize the orbitals by a unitary rotation eκe^{\boldsymbol{\kappa}} and minimize the energy functional directly with respect to the anti-Hermitian matrix κ\boldsymbol{\kappa} using its gradient and an exact or approximate Hessian. [85, 86] Because every step lowers the energy, these methods converge reliably to a minimum, which makes them robust for ground states. Δ\DeltaSCF excited states, however, are typically saddle points of the energy and can collapse to lower solutions unless the optimization is modified. [20, 30] Throughout this work, we generate static SCF reference data with both standard DIIS and the second-order co-iterative augmented-Hessian (CIAH) orbital optimizer, [87] as implemented in PySCF. [88, 89]

Neither strategy completely removes the difficulties that static SCF poses in a molecular-dynamics setting. Converging a full SCF calculation at every time step is expensive, and convergence itself becomes unreliable in the strong-coupling regions that govern nonadiabatic transitions. In particular, near avoided crossings, where orbital energies are close to degenerate, the SCF iteration may oscillate between competing configurations or fail outright, producing discontinuous energies and forces.

The Ground and Δ\DeltaSCF Excited States

In this work, we restrict ourselves to two states: the ground state (N=0N = 0) and the lowest singly excited singlet (N=1N = 1). We define them here in the static picture, that is, as stationary SCF solutions converged at a fixed nuclear geometry, before turning to their dynamical propagation. Both are described in a spin-restricted formalism throughout. Applications involving bond breaking or pronounced open-shell character will require an extension to unrestricted Kohn–Sham theory, which we leave to future work.

  • Ground state (N=0N = 0). The standard Aufbau solution, in which the lowest orbitals are doubly occupied (fn=2f_n = 2) and the remaining orbitals are empty (fn=0f_n = 0), as depicted in Figure 1a. The many-electron Kohn–Sham (KS) wavefunction is the single Slater determinant

Figure 1. Molecular-orbital occupation patterns of (a) the ground state (N=0N = 0) and (b) the first excited state (N=1N = 1).

Φ0=∣ψ1ψˉ1…ψHOMOψˉHOMO⟩(6)\Phi_0 = \lvert\psi_1 \bar{\psi}_1 \ldots\psi_{\mathrm{HOMO}} \bar{\psi}_{\mathrm{HOMO}} \rangle \tag*{(6)}
  • Excited state (N=1N = 1). A non-Aufbau configuration in which one electron is promoted from the HOMO to the LUMO (Figure 1b). Formally, the spin-restricted open-shell singlet is the two-determinant combination

Φ1=12(∣ψ1ψˉ1…ψHOMOψˉLUMO⟩−∣ψ1ψˉ1…ψˉHOMOψLUMO⟩)(7)\Phi_1 = \frac{1}{\sqrt{2}}\left(\lvert\psi_1 \bar{\psi}_1 \ldots\psi_{\mathrm{HOMO}} \bar{\psi}_{\mathrm{LUMO}} \rangle- \lvert\psi_1 \bar{\psi}_1 \ldots\bar{\psi}_{\mathrm{HOMO}} \psi_{\mathrm{LUMO}} \rangle\right) \tag*{(7)}

However, we work throughout with a spin-restricted, occupation-constrained Δ\DeltaSCF description,[90] in which a single set of restricted spatial orbitals is used, with occupation numbers fn=1f_n = 1 for the two singly occupied orbitals (HOMO and LUMO) and fn=2f_n = 2 for the remaining occupied orbitals. The energies and the corresponding nuclear forces are therefore all obtained from the spin-restricted KS energy functional evaluated on the density of Eq. (3) built with these occupation numbers, that is, with a spin-unpolarized density. This energy differs both from that of the spin-resolved mixed determinant and from the spin-purified singlet energy of ROKS,[31, 32] and we use it consistently in the static and dynamical calculations. Equation 7 is used only to define the spin-adapted many-electron wavefunction Φ1\Phi_1 entering the state overlaps of § 3.3, where it determines the form of Eq. (32).

The adiabatic states are calculated differently in the static and dynamical settings, and the distinction matters for the comparisons made later. For the static reference calculations, the maximum-overlap method (MOM)[27] is applied at every geometry along the trajectory. In the dynamical scheme, by contrast, MOM is used only once, to converge the state at the initial time, and thereafter the excited-state character is maintained by the energy-scaled propagation itself, through the continuity of the time evolution. Dispensing with the orbitalselection criterion at every step is one of the practical advantages of the present approach, and, as we show later, it helps the propagated trajectory stay on the same branch even when the static solver jumps to a different root (§ 4.2).

Adiabatic-State Following via Real-Time Propagation

The convergence problems near degeneracies stem from treating each geometry as an independent variational problem. We therefore replace repeated static optimizations by a dynamical propagation scheme, in which the target state is carried from one geometry to the next by the continuity of its own time evolution. This scheme rests on the quantum adiabatic theorem.

The Quantum Adiabatic Theorem

The quantum adiabatic theorem states that a system initialized in an eigenstate of a time-dependent Hamiltonian H^(t)\hat{H}(t) remains in the corresponding instantaneous eigenstate ΦN(t)\Phi_N(t) provided the Hamiltonian changes slowly on the characteristic electronic timescale. [91, 92, 93] In other words, the timescale of the external perturbation, here driven by the nuclear motion, must be long compared with the electronic response time τel∼1/ΔE\tau_{\mathrm{el}} \sim1/\Delta E, where ΔE\Delta E is the relevant energy gap. To be more quantitative, one can use Landau–Zener theory to provide estimates of whether a transition will be adiabatic given the velocity of the nuclei. [94, 95]

When this separation of timescales holds, RT-TDDFT coupled to Ehrenfest dynamics follows an approximately adiabatic trajectory, much as Born–Oppenheimer molecular dynamics does. [96] However, at high nuclear velocities or for small energy gaps, as near an avoided crossing, the adiabatic condition fails, and the electronic wavefunction develops into a superposition of adiabatic states. The forces resulting from such a state are a mean field over the adiabatic states and thus lead to the Ehrenfest description of nonadiabatic dynamics.

In order to go beyond the mean-field description, surface hopping requires pure adiabatic states with state-specific forces. Standard RT-TDDFT is therefore not directly suited to generating the underlying adiabatic surfaces. Within the time-dependent formalism, we therefore fore introduce an energy-scaling parameter α>1\alpha> 1 through the transformation H^KS↦αH^KS\hat{H}_{\mathrm{KS}} \mapsto\alpha\hat{H}_{\mathrm{KS}}, which raises the frequency of the electronic phase evolution relative to the nuclear motion and so widens the separation of the two timescales and recovers the adiabatic condition.

Energy-Scaled Equations of Motion

In this section, we focus on the electronic dynamics and leave the formulation of the coupled electron–nuclear nonadiabatic dynamics to § 3. We therefore consider a prescribed nuclear trajectory R(t)\mathbf{R}(t), for which we define the electronic action over the interval [t0,t1][t_0,t_1],

A=∫t0t1dt{∑nfn⟨ψn(t)|i∂∂t|ψn(t)⟩−αEKS[ρ(t);R(t)]}(8)\mathcal{A} = \int_{t_0}^{t_1} \mathrm{d}t \left\{ \sum_n f_n \left\langle\psi_n(t) \middle| \mathrm{i}\frac{\partial}{\partial t} \middle| \psi_n(t) \right\rangle- \alpha E_{\mathrm{KS}}[\rho(t);\mathbf{R}(t)] \right\} \tag*{(8)}

where EKSE_{\mathrm{KS}} is the functional of Eq. (2) evaluated on the instantaneous density ρ(t)\rho(t), built from the propagated TDKS orbitals with fixed occupation numbers. The initial orbitals {ψn(t0)}\{\psi_n(t_0)\} are taken to be a stationary converged solution of Eq. (5). Making this action stationary yields the energy-scaled TDKS equations

i∂∂t∣ψn(t)⟩=αH^KS(t)∣ψn(t)⟩(9)\mathrm{i}\frac{\partial}{\partial t}|\psi_n(t)\rangle= \alpha\hat{H}_{\mathrm{KS}}(t)|\psi_n(t)\rangle \tag*{(9)}

with H^KS(t)≡H^KS[ρ(t);R(t)]\hat{H}_{\mathrm{KS}}(t) \equiv\hat{H}_{\mathrm{KS}}[\rho(t);\mathbf{R}(t)].

Setting α=1\alpha= 1 recovers the standard real-time TDKS equations. The introduction of the scaling parameter α\alpha can be interpreted as a change of the electronic clock. Defining the rescaled time s=αts = \alpha t, Eq. (9) becomes

i∂∂s∣ψn(s)⟩=H^KS[ρ(s);R(s/α)]∣ψn(s)⟩(10)\mathrm{i}\frac{\partial}{\partial s}|\psi_n(s)\rangle= \hat{H}_{\mathrm{KS}}[\rho(s);\mathbf{R}(s/\alpha)]|\psi_n(s)\rangle \tag*{(10)}

which is the standard, unscaled TDKS equation propagated along the same nuclear path traversed at 1/α1/\alpha of the original speed. Scaling the electronic Hamiltonian by α\alpha is therefore equivalent to slowing the nuclei by the same factor. The equivalence is exact and survives the nonlinearity of H^KS\hat{H}_{\mathrm{KS}} because ρ\rho is built from the instantaneous TDKS orbitals and is reparametrized along with them. Note that the scaling enters only the orbital equation of motion. All energies and nuclear forces, including the potential energies entering MASH, are evaluated from the unscaled KS energy functional on the propagated density, so the potential-energy surfaces themselves are unchanged.

Through Eq. (10), the adiabatic theorem of § 2.3.1 suggests that in the α→∞\alpha\to\infty limit, where the electrons experience an infinitely slow nuclear motion, each propagated state rigorously recovers the instantaneous adiabatic state selected by its initial condition, provided this state stays separated from the others by a finite gap. However, if the gap is zero, it is clear that no finite value of α\alpha can be found which obeys the adiabatic principle. Such a case is called a trivial crossing and there is nonetheless a simple remedy, due to the fact that the propagated states will retain their electronic character while exchanging their energy ordering. Thus the correct ordering can be easily restored by state tracking and surface switching.[97, 98, 99] No trivial crossings occurred in the trajectories reported in this work.

We note that the electronic equation of motion, Eq. (9), is mathematically equivalent to that of the modified Ehrenfest dynamics of Rubio and coworkers[45, 46], iμ∂tψn=H^KSψni\mu\partial_t\psi_n = \hat{H}_{\mathrm{KS}}\psi_n, with α=1/μ\alpha= 1/\mu. The two methods, however, use this parameter for different purposes. Modified Ehrenfest dynamics was designed as an efficient alternative to ground-state Born–Oppenheimer molecular dynamics and takes μ≫1\mu\gg1, that is, α≪1\alpha\ll1. The electronic excitation energies are then scaled down by 1/μ1/\mu and the time step can be increased by a factor of μ\mu, at the price of weaker adiabaticity, so the admissible μ\mu is limited by the electronic gap. Those studies also discussed the adiabatic limit at small μ\mu. Here we work at α>1\alpha> 1 and accept a correspondingly smaller time step in exchange for adiabatic following, with α\alpha chosen large enough to obtain numerical convergence. As demonstrated in § 4.2, this lets us follow Δ\DeltaSCF excited states as well as the ground state through regions where static SCF fails, thereby providing the adiabatic states and couplings that surface hopping requires. In practice, Eq. (9) differs from the standard RT-TDDFT equation only by the scalar factor α\alpha, so the energy scaling itself requires only a minor modification of the orbital propagation in an existing RT-TDDFT Ehrenfest code.

Numerical Implementation in an Atom-Centered Orbital Basis

We represent the time-dependent KS orbitals in a basis of atomic orbitals (AOs) using the linear-combination-of-atomic-orbitals (LCAO) expansion

ψn(r,t)=∑μcμn(t)ϕμ(r−RA(t))(11)\psi_n(\mathbf{r},t)=\sum_{\mu}c_{\mu n}(t)\phi_{\mu}\left(\mathbf{r}-\mathbf{R}_A(t)\right) \tag*{(11)}

such that each basis function ϕμ\phi_{\mu} is centered on its corresponding nucleus AA at its instantaneous position RA(t)\mathbf{R}_A(t). The basis functions thus acquire an implicit time dependence through the nuclear motion. In this way, the LCAO expansion employs a moving diabatic basis of local functions attached to their nuclei.

Substituting the expansion into Eq. (9) and projecting onto the atomic basis gives the matrix equation of motion for the coefficients. Because the basis functions are themselves time dependent, the chain rule applied to ∂tψn\partial_t\psi_n produces two terms, ∑μ(c˙μnϕμ+cμnϕ˙μ)\sum_{\mu}\left(\dot{c}_{\mu n}\phi_{\mu}+c_{\mu n}\dot{\phi}_{\mu}\right), the second of which has no counterpart in a fixed basis. Projecting onto ϕμ\phi_{\mu} we obtain

∑νSμνc˙νn=−∑ν(iαHμν+Dμν)cνn(12)\sum_{\nu}S_{\mu\nu}\dot{c}_{\nu n}=-\sum_{\nu}\left(\mathrm{i}\alpha H_{\mu\nu}+D_{\mu\nu}\right)c_{\nu n} \tag*{(12)}

where the matrix elements are

Sμν=⟨ϕμ(t)∣ϕν(t)⟩(13)S_{\mu\nu}=\langle\phi_{\mu}(t)|\phi_{\nu}(t)\rangle \tag*{(13)}
Hμν=⟨ϕμ(t)∣H^KS(t)∣ϕν(t)⟩(14)H_{\mu\nu}=\langle\phi_{\mu}(t)|\hat{H}_{\mathrm{KS}}(t)|\phi_{\nu}(t)\rangle \tag*{(14)}
Dμν=⟨ϕμ(t)∣ϕ˙ν(t)⟩=∑AR˙A(t)⋅⟨ϕμ(R(t))∣∇Aϕν(R(t))⟩(15)D_{\mu\nu}=\langle\phi_{\mu}(t)|\dot{\phi}_{\nu}(t)\rangle=\sum_A\dot{\mathbf{R}}_A(t)\cdot\langle\phi_{\mu}(\mathbf{R}(t))|\nabla_A\phi_{\nu}(\mathbf{R}(t))\rangle \tag*{(15)}

Here S and H are the AO overlap and KS Hamiltonian matrices, and D is the time-derivative of the basis matrix. coupling induced by the motion of the basis functions, expressed in Eq. 13c in terms of the nuclear velocities R˙A\dot{\mathbf{R}}_A by the chain rule, with ∇A\nabla_A the gradient with respect to the position coordinate RA\mathbf{R}_A of nucleus A. Because the basis depends on time only through the nuclear positions, we also write S(t)≡S(R(t))\mathbf{S}(t) \equiv\mathbf{S}(\mathbf{R}(t)) below. The elements ⟨ϕμ∣∇Aϕν⟩\langle\phi_\mu\mid\nabla_A \phi_\nu\rangle have a form similar to that of the nonadiabatic coupling vectors between many-electron states, but here describe purely geometric basis transport. In particular, they account for the change in the orbital expansion as the basis functions move with the nuclei.

Equation 12 cannot be integrated directly, for two reasons. First, the Hamiltonian H[ρ(t)]\mathbf{H}[\rho(t)] is nonlinear through its dependence on the instantaneous density and hence on the coefficients themselves. Second, the matrices D\mathbf{D} and S\mathbf{S} carry the geometric dependence of a basis that is itself in motion, which makes stable, unitary propagation delicate.[100]

We therefore apply a second-order symmetric Trotter (Strang) splitting of the short-time propagator

U^(Δt)≃U^orb(Δt2)U^coeff(Δt)U^orb(Δt2)(16)\hat{U}(\Delta t) \simeq\hat{U}_{\mathrm{orb}}\left(\frac{\Delta t}{2}\right)\hat{U}_{\mathrm{coeff}}(\Delta t)\hat{U}_{\mathrm{orb}}\left(\frac{\Delta t}{2}\right) \tag*{(16)}

Here U^orb\hat{U}_{\mathrm{orb}} transports the orbitals along the trajectory from one geometry to another time Δt/2\Delta t/2 later, whereas U^coeff\hat{U}_{\mathrm{coeff}} evolves the TDKS coefficients forward by time Δt\Delta t at the fixed midpoint geometry. Splitting the propagator this way replaces the coupled evolution by two sub-problems that can each be integrated accurately.

The first operation, U^orb\hat{U}_{\mathrm{orb}}, corresponds to the D\mathbf{D} term of Eq. (12). Rather than evaluating the derivative-coupling matrix explicitly,[52] we use the numerically more stable Löwdin-orthonormalized transport scheme,[100, 101] which maps coefficients expressed in the basis at geometry R\mathbf{R} onto the basis at R′\mathbf{R}',

C~=S(R′)−1/2S(R)1/2C(17)\widetilde{\mathbf{C}} = \mathbf{S}(\mathbf{R}')^{-1/2}\mathbf{S}(\mathbf{R})^{1/2}\mathbf{C} \tag*{(17)}

where C\mathbf{C} is a matrix of AO coefficients cμnc_{\mu n} of the TDKS orbitals and C~\widetilde{\mathbf{C}} contains the corresponding coefficients in the basis at R′\mathbf{R}'. The transport is applied twice per step: first, from R(t)\mathbf{R}(t) to R(t+Δt/2)\mathbf{R}(t+\Delta t/2), which turns C(t)\mathbf{C}(t) into C~(t)\widetilde{\mathbf{C}}(t); then, after the coefficients are updated from C~(t)\widetilde{\mathbf{C}}(t) to C~(t+Δt)=U^coeffC~(t)\widetilde{\mathbf{C}}(t+\Delta t)=\widehat{U}_{\mathrm{coeff}}\widetilde{\mathbf{C}}(t), the basis is transported from R(t+Δt/2)\mathbf{R}(t+\Delta t/2) to R(t+Δt)\mathbf{R}(t+\Delta t), which turns C~(t+Δt)\widetilde{\mathbf{C}}(t+\Delta t) into C(t+Δt)\mathbf{C}(t+\Delta t). Because the transformation is a product of matrix square roots of overlap matrices, it preserves orthonormality exactly: C~†S(R′)C~=C†S(R)C=I\widetilde{\mathbf{C}}^\dagger\mathbf{S}(\mathbf{R}^{\prime})\widetilde{\mathbf{C}}=\mathbf{C}^\dagger\mathbf{S}(\mathbf{R})\mathbf{C}=\mathbf{I} (using the fact that S\mathbf{S} is a Hermitian matrix). It is nevertheless an approximation to the exact action of the D\mathbf{D} term: it neglects the part of the basis motion that is not reflected in the change of the overlap matrix, for example a rigid translation of the molecule, which leaves S\mathbf{S} unchanged.[100] A projection using the mixed overlap ⟨ϕμ(R′)∣ϕν(R)⟩\langle\phi_{\mu}(\mathbf{R}^{\prime})|\phi_{\nu}(\mathbf{R})\rangle would retain this part but does not in general preserve orthonormality in a finite basis.

The second operation, U^coeff\widehat{U}_{\mathrm{coeff}}, evolves the electronic coefficients and corresponds to the H\mathbf{H} term of Eq. (12). In the transported basis, the coefficients then obey

iS~C~˙=αH[ρ]C~(18)i\widetilde{\mathbf{S}}\dot{\widetilde{\mathbf{C}}}=\alpha\mathbf{H}[\rho]\widetilde{\mathbf{C}} \tag*{(18)}

with S~≡S(R(t+Δt/2))\widetilde{\mathbf{S}}\equiv\mathbf{S}(\mathbf{R}(t+\Delta t/2)) and with H\mathbf{H} constructed at the same geometry. We integrate this equation with the implicit Crank–Nicolson scheme,[102, 103] chosen because it is unitary and time reversible,

[S~+iαΔt2H‾]C~(t+Δt)=[S~−iαΔt2H‾]C~(t)(19)\left[\widetilde{\mathbf{S}}+i\alpha\frac{\Delta t}{2}\overline{\mathbf{H}}\right]\widetilde{\mathbf{C}}(t+\Delta t)=\left[\widetilde{\mathbf{S}}-i\alpha\frac{\Delta t}{2}\overline{\mathbf{H}}\right]\widetilde{\mathbf{C}}(t) \tag*{(19)}

where the time-averaged Hamiltonian is

H‾=12(H[ρ(t)]+H[ρ(t+Δt)])(20)\overline{\mathbf{H}}=\frac{1}{2}\left(\mathbf{H}[\rho(t)]+\mathbf{H}[\rho(t+\Delta t)]\right) \tag*{(20)}

Because H[ρ(t+Δt)]\mathbf{H}[\rho(t+\Delta t)] depends on a final density that is not known beforehand, Eq. (19) is nonlinear and must be solved self-consistently. We initialize the time-averaged Hamiltonian from its value at the start of the step, solve Eq. (19), rebuild H‾\overline{\mathbf{H}} from the resulting density, and repeat.[104] Without acceleration, this fixed-point iteration converges slowly in the stiff, large-α\alpha regime, so we accelerate it with Anderson mixing[105] over the history of Hamiltonian matrices, which constructs each new input Hamiltonian as the linear combination of previous input and output Hamiltonians that minimizes the norm of their residual, and stop when the largest change in the molecular-orbital coefficients between successive iterations falls below the convergence threshold. Because consecutive geometries along the trajectory are close, the initial guess is good and convergence is typically reached within a few iterations.

The nuclear force for each state NN is the analytic gradient of its potential energy VNV_N, evaluated with the instantaneous propagated density, as in Ehrenfest dynamics. Writing P=∑nfncncn†\mathbf{P} = \sum_n f_n \mathbf{c}_n \mathbf{c}_n^\dagger for the one-particle density matrix in the atomic-orbital basis, built from the columns cn\mathbf{c}_n of C\mathbf{C} with the occupations of Sec. 2.2, the force on nucleus AA for this state is

FA=−∂VN∂RA∣C+Tr⁡[∂S∂RA12(S−1HP+PHS−1)](21)\mathbf{F}_A = -\left.\frac{\partial V_N}{\partial\mathbf{R}_A}\right|_{\mathbf{C}} + \operatorname{Tr}\left[\frac{\partial\mathbf{S}}{\partial\mathbf{R}_A}\frac{1}{2}\left(\mathbf{S}^{-1}\mathbf{H}\mathbf{P}+\mathbf{P}\mathbf{H}\mathbf{S}^{-1}\right)\right] \tag*{(21)}

The first term is the explicit derivative of VNV_N at fixed AO coefficients. It contains the Hellmann–Feynman force,[106] including the nuclear repulsion, together with derivatives of the AO integrals that arise because the basis functions travel with the nuclei. The second term comes from the derivative of the overlap matrix; together with these integral derivatives it constitutes the Pulay force.[107] In Born–Oppenheimer dynamics this term involves the energy-weighted density matrix[69] W=∑nfnεncncn†\mathbf{W} = \sum_n f_n \varepsilon_n \mathbf{c}_n \mathbf{c}_n^\dagger, whose very definition presupposes that the orbitals are eigenstates of the Kohn–Sham Hamiltonian with eigenvalues εn\varepsilon_n. Propagated TDKS orbitals are not in general eigenstates of H^KS\hat{H}_{\mathrm{KS}}, so no such eigenvalues exist, and 12(S−1HP+PHS−1)\frac{1}{2}(\mathbf{S}^{-1}\mathbf{H}\mathbf{P}+\mathbf{P}\mathbf{H}\mathbf{S}^{-1}) takes the place of W\mathbf{W};[2, 51, 108] the two coincide whenever the orbitals do satisfy the stationary Kohn–Sham equations, so Eq. (21) reduces to the familiar Born–Oppenheimer gradient in that limit. Following common practice,[2, 51] we omit the further terms involving the velocities of the basis functions, which contribute to momentum conservation but not to the energy gradient.

Algorithm 1 summarizes the propagation scheme. It applies equally to the ground- and the excited-state TDKS orbitals: each is initialized from a converged static SCF solution at the starting geometry and then propagated independently along the nuclear trajectory R(t)\mathbf{R}(t). Provided α\alpha is large enough for the propagated states to follow their adiabatic states (§ 2.3), each propagated state reproduces the energy and force of the corresponding adiabatic potential-energy surface on the fly, with neither repeated diagonalization nor explicit optimization of the energy functional at every step. This already allows Born–Oppenheimer dynamics on either the ground or the excited state. To treat nonadiabatic processes we still need a surface-hopping scheme that determines which state drives the nuclei, together with the couplings between the two states on which that scheme relies. This is the subject of the next section.

Algorithm 1 Propagation of a single TDKS state over one step, from R(t)\mathbf{R}(t) to R(t+Δt)\mathbf{R}(t+\Delta t).

input: coefficients C≡C(t)\mathbf{C} \equiv\mathbf{C}(t) at R(t)\mathbf{R}(t), scaling α\alpha, step Δt\Delta t; S~≡S(t+Δt/2)\widetilde{\mathbf{S}} \equiv\mathbf{S}(t+\Delta t/2) transport to the midpoint basis: C~←S~−1/2S(t)1/2C\widetilde{\mathbf{C}} \leftarrow\widetilde{\mathbf{S}}^{-1/2}\mathbf{S}(t)^{1/2}\mathbf{C} initialize H′←H[ρ(C~)]\mathbf{H}' \leftarrow\mathbf{H}[\rho(\widetilde{\mathbf{C}})] and H‾←H′\overline{\mathbf{H}} \leftarrow\mathbf{H}' at R(t+Δt/2)\mathbf{R}(t+\Delta t/2) repeat     solve (S~+iαΔt H‾/2)C~′=(S~−iαΔt H‾/2)C~(\widetilde{\mathbf{S}}+i\alpha\Delta t\,\overline{\mathbf{H}}/2)\widetilde{\mathbf{C}}'=(\widetilde{\mathbf{S}}-i\alpha\Delta t\,\overline{\mathbf{H}}/2)\widetilde{\mathbf{C}} for C~′\widetilde{\mathbf{C}}'     update H′\mathbf{H}' by Anderson mixing with H[ρ(C~′)]\mathbf{H}[\rho(\widetilde{\mathbf{C}}')]; H‾←(H[ρ(C~)]+H′)/2\overline{\mathbf{H}}\leftarrow(\mathbf{H}[\rho(\widetilde{\mathbf{C}})]+\mathbf{H}')/2 until the change in C~′\widetilde{\mathbf{C}}' is below the threshold transport to the end-point basis: C′←S(t+Δt)−1/2S~1/2C~′\mathbf{C}'\leftarrow\mathbf{S}(t+\Delta t)^{-1/2}\widetilde{\mathbf{S}}^{1/2}\widetilde{\mathbf{C}}' return C′≡C(t+Δt)\mathbf{C}'\equiv\mathbf{C}(t+\Delta t), with energy and forces at R(t+Δt)\mathbf{R}(t+\Delta t) evaluated from ρ(C′)\rho(\mathbf{C}')

Surface Hopping in a Complex Adiabatic Basis

To carry out nonadiabatic molecular dynamics, we combine the on-the-fly real-time propagation scheme developed in § 2 with the mapping approach to surface hopping (MASH), [6, 7] which provides equations of motion for the time-dependent coefficients of the two adiabatic electronic states and a rule for when the nuclei switch surface. The standard MASH formulation shares much of its algorithmic structure with FSSH but is deterministic and time reversible.[109] It is free of the internal-consistency problems of FSSH,[110, 111] correctly obeys detailed balance,[112] and can be used to calculate reaction rates,[113] sample rare events,[114, 115] capture electronic coherences, [116, 117] describe open quantum systems, [118] and perform on-the-fly simulations of photochemistry. [119, 120, 121]

Although MASH was originally formulated for two states, various multistate generalizations have been proposed. [123, 124] [122] Using MASH in the present framework requires two extensions. First, previous surface-hopping implementations based on static electronic-structure calculations assume real-valued wavefunctions, [17, 77, 82, 109, 119, 120] [23] whereas real-time propagation necessarily yields complex-valued states carrying time-dependent phases; we therefore recast the MASH equations of motion in a complex adiabatic basis. Second, the ground and excited states are obtained from different SCF solutions, and so do not share a common KS Hamiltonian and are not mutually orthogonal; we therefore evaluate the time-derivative couplings and nonadiabatic coupling vectors numerically between the two independently propagated adiabatic states.

MASH Formulation for Complex Adiabatic States

First, we assume that the propagated adiabatic states have been orthogonalized using the procedure that we will describe in § 3.2. The time-dependent electronic wavefunction can then be written as ∣ψ(t)⟩=c0(t)∣Φ0(R(t))⟩+c1(t)∣Φ1(R(t))⟩|\psi(t)\rangle= c_0(t)|\Phi_0(\boldsymbol{R}(t))\rangle+ c_1(t)|\Phi_1(\boldsymbol{R}(t))\rangle, so that c0c_0 and c1c_1 describe the amplitude on each state. In MASH, this two-state amplitude is mapped onto a classical spin vector S⃗={Sx,Sy,Sz}\vec{S} = \{S_x, S_y, S_z\} on the unit Bloch sphere, whose Cartesian components are the expectation values of the Pauli operators, Sμ=⟨σ^μ⟩S_\mu= \langle\hat{\sigma}_\mu\rangle.[6] In the two-state adiabatic basis {∣Φ0⟩,∣Φ1⟩}\{|\Phi_0\rangle, |\Phi_1\rangle\} these operators read

σ^x=∣Φ0⟩⟨Φ1∣+∣Φ1⟩⟨Φ0∣(22)\hat{\sigma}_x = |\Phi_0\rangle\langle\Phi_1| + |\Phi_1\rangle\langle\Phi_0| \tag*{(22)}
σ^y=i(∣Φ0⟩⟨Φ1∣−∣Φ1⟩⟨Φ0∣)(23)\hat{\sigma}_y = i\left(|\Phi_0\rangle\langle\Phi_1| - |\Phi_1\rangle\langle\Phi_0|\right) \tag*{(23)}
σ^z=∣Φ1⟩⟨Φ1∣−∣Φ0⟩⟨Φ0∣(24)\hat{\sigma}_z = |\Phi_1\rangle\langle\Phi_1| - |\Phi_0\rangle\langle\Phi_0| \tag*{(24)}

Hence, the mapping variables are Sx=2Re⁡(c0∗c1)S_x = 2\operatorname{Re}(c_0^*c_1), Sy=−2Im⁡(c0∗c1)S_y = -2\operatorname{Im}(c_0^*c_1) and Sz=∣c1∣2−∣c0∣2S_z = |c_1|^2 - |c_0|^2.

To obtain the spin dynamics, we first write the equations of motion for the coefficients in the complex adiabatic basis. KS-DFT does not provide a universal many-electron Hamiltonian common to these state-specific solutions, so we instead define an effective diagonal Hamiltonian whose eigenvalues are the adiabatic potential-energy surfaces {VN(R)}\{V_N(\mathbf{R})\}, where VN(R)V_N(\mathbf{R}) denotes the unscaled potential energy V(R)V(\mathbf{R}) of Sec. 2.1, evaluated for state NN with the occupations specified in Sec. 2.2. The states are coupled by the time-derivative coupling matrix TMN=⟨ΦM∣∂tΦN⟩T_{MN} = \langle\Phi_M|\partial_t\Phi_N\rangle, which is anti-Hermitian, with purely imaginary diagonal elements that need not vanish. In this basis, the coefficients {c0,c1}\{c_0,c_1\} obey the time-dependent Schrödinger equation,

c˙M=−iVMcM−∑NTMNcN(25)\dot{c}_M = -iV_Mc_M - \sum_N T_{MN}c_N \tag*{(25)}

which, together with TNN∗=−TNNT_{NN}^{*}=-T_{NN} and T10=−T01∗T_{10}=-T_{01}^{*}, gives the equation of motion of the spin vector on the Bloch sphere,

(S˙xS˙yS˙z)=(0−Ω−2Re⁡(T01)Ω0−2Im⁡(T01)2Re⁡(T01)2Im⁡(T01)0)(SxSySz)(26)\begin{pmatrix} \dot{S}_x \\ \dot{S}_y \\ \dot{S}_z \end{pmatrix} = \begin{pmatrix} 0 & -\Omega& -2\operatorname{Re}(T_{01}) \\ \Omega& 0 & -2\operatorname{Im}(T_{01}) \\ 2\operatorname{Re}(T_{01}) & 2\operatorname{Im}(T_{01}) & 0 \end{pmatrix} \begin{pmatrix} S_x \\ S_y \\ S_z \tag*{(26)} \end{pmatrix}

where Ω=ΔV−Im⁡(T00−T11)\Omega= \Delta V - \operatorname{Im}(T_{00}-T_{11}) and ΔV=V1−V0\Delta V = V_1 - V_0.

The diagonal couplings TNNT_{NN} vanish for real-valued adiabatic states, whereas in a complex basis they account for the phase evolution of the basis states. In Eq. (26), they enter through Ω\Omega, which sets the precession rate about the zz axis in the chosen phase convention. For a model Hamiltonian with eigenenergies EN=VNE_N = V_N, the natural phase evolution of ∣ΦN(t)⟩=e−iENt∣ΦN⟩|\Phi_N(t)\rangle= e^{-iE_Nt}|\Phi_N\rangle gives TNN=−iENT_{NN}=-iE_N and Ω=0\Omega=0. The basis states then carry the dynamical phases, and the coefficients are expressed in a co-rotating frame.

However, under real-time KS propagation, the phases carried by the basis states need not be generated by the total energies VNV_N. For example, at fixed nuclear geometry and before any additional gauge fixing, a determinant built from stationary orbitals acquires a phase governed by the sum of the occupied orbital energies multiplied by α\alpha from Eq. (9). In KS-DFT, the sum of occupied orbital energies generally differs from the total energy VNV_N. In this case, the diagonal terms TNNT_{NN} account for the phase evolution of the basis in Eq. (25) and cannot be set to zero.

The dynamics is formally independent of the phase convention due to gauge invariance. To make this explicit, consider a change of the basis functions by an arbitrary real phase θN(t)\theta_N(t), accompanied by a corresponding change of the expansion coefficients,

∣ΦN′⟩=eiθN(t)∣ΦN⟩,cN′=e−iθN(t)cN,\lvert\Phi'_N\rangle= e^{i\theta_N(t)}\lvert\Phi_N\rangle,\qquad c'_N=e^{-i\theta_N(t)}c_N,

such that the total electronic wavefunction is unchanged. Under this transformation, T01T_{01} and Sx−iSyS_x-iS_y acquire opposite phase factors, leaving the population equation

S˙z=2Re⁡(T01)Sx+2Im⁡(T01)Sy=2Re⁡[T01(Sx−iSy)](27)\dot{S}_z=2\operatorname{Re}(T_{01})S_x+2\operatorname{Im}(T_{01})S_y=2\operatorname{Re}\left[T_{01}\left(S_x-iS_y\right)\right] \tag*{(27)}

unchanged. The momentum-rescaling combination Re⁡[d01(Sx−iSy)]\operatorname{Re}\left[\mathbf{d}_{01}\left(S_x-iS_y\right)\right] in Eq. (28) is invariant for the same reason. Thus, populations, hopping events and rescaling directions are independent of the phase convention used for the basis states.

The population equation also influences the nuclear dynamics, since crossing the Bloch-sphere equator at Sz=0S_z=0 triggers an attempt to switch the active surface. The nuclear equations of motion follow from requiring conservation of the total MASH energy,1 E=∑APA2/2MA+h(−Sz)V0+h(Sz)V1\mathcal{E}=\sum_A\mathbf{P}_A^2/2M_A+h(-S_z)V_0+h(S_z)V_1, where hh is the Heaviside step function and MAM_A is the mass of atom AA. Setting dE/dt=0d\mathcal{E}/dt=0, substituting S˙z\dot{S}_z from Eq. (27) and using R˙A=PA/MA\dot{\mathbf{R}}_A=\mathbf{P}_A/M_A and T01=∑AR˙A⋅d01AT_{01}=\sum_A\dot{\mathbf{R}}_A\cdot\mathbf{d}_{01}^A, where d01A=⟨Φ0∣∇AΦ1⟩\mathbf{d}_{01}^A=\langle\Phi_0\lvert\nabla_A\Phi_1\rangle is the nonadiabatic coupling vector (NACV) between the two propagated states, gives the generalized nuclear force

P˙A=−∇AV0h(−Sz)−∇AV1h(Sz)−2ΔVδ(Sz)[Re⁡(d01A)Sx+Im⁡(d01A)Sy](28)\dot{\mathbf{P}}_A=-\nabla_A V_0h(-S_z)-\nabla_A V_1h(S_z)-2\Delta V\delta(S_z)\left[\operatorname{Re}(\mathbf{d}_{01}^A)S_x+\operatorname{Im}(\mathbf{d}_{01}^A)S_y\right] \tag*{(28)}

The impulsive term at Sz=0S_z = 0 involves the NACV d01\mathbf{d}_{01}. For complex-valued states, only the direction of the combination Re⁡(d01A)Sx+Im⁡(d01A)Sy\operatorname{Re}(\mathbf{d}_{01}^{\mathrm{A}})S_x + \operatorname{Im}(\mathbf{d}_{01}^{\mathrm{A}})S_y enters the momentum rescaling. Each time a hop is attempted, the kinetic energy available along that direction is compared with the energy gap to decide whether the hop is energetically allowed; if it is, the momentum is rescaled along that direction, and if not the hop is frustrated. Standard MASH handles a frustrated hop by reversing the momentum along the NACV.[6, 7] We instead use the recently proposed π\pi-MASH prescription to flip the spin vector, S⃗←−S⃗\vec{S} \leftarrow-\vec{S}, and leave the momentum unchanged.[125]

DTO Representation of the Adiabatic States

Evaluating the time-derivative coupling matrix T\mathbf{T} and the NACV d01\mathbf{d}_{01} requires a well-defined representation of the adiabatic states. Within the KS-DFT framework, constructing many-electron wavefunctions from KS Slater determinants is a practical approximation routinely used in static Δ\DeltaSCF.[77] [23] Even in static Δ\DeltaSCF, the ground and excited states are optimized independently, so their mutual orthogonality is not enforced. Real-time propagation introduces an additional ambiguity: the two equally occupied active spatial orbitals of the excited state may mix unitarily without changing the KS density, while changing the spin-adapted excited-state wavefunction constructed from them. In the strong-coupling region, this mixing can give both active orbitals substantial components in the occupied ground-state space. The singlet constructed from them can then develop a large overlap with the ground-state determinant (§ 4.3, Figure 3), compromising the overlap-based evaluation of the couplings. This behavior resembles the orbital-rotation instability found in ROKS simulations.[79, 80]

Figure 3. Magnitude of the overlap between the ground- and excited-state wavefunctions, ∣⟨Φ0(t)∣Φ1(t)⟩∣|\langle\Phi_0(t)|\Phi_1(t)\rangle|, propagated with α=10\alpha= 10 along the RT-TDDFT/Ehrenfest path of Figure 2(b). Wavefunctions constructed directly from the propagated TDKS orbitals (blue solid curve) are compared with those obtained in the DTO gauge (orange dashed curve).

We fix the particle–hole representation by transforming the active TDKS orbitals into dynamical transition orbitals (DTOs).[81] The construction uses spin-restricted orbitals with equal occupations in the active space. We first build the projector onto the occupied groundstate (N=0)(N = 0) orbitals,

P^0(t)=∑n∈occ∣ψn(0)(t)⟩⟨ψn(0)(t)∣(29)\hat{P}_0(t) = \sum_{n \in\mathrm{occ}} \lvert\psi_n^{(0)}(t) \rangle\langle\psi_n^{(0)}(t) \rvert \tag*{(29)}

and then the projection matrix M(t)\mathbf{M}(t) within the subspace spanned by the two time-dependent singly occupied orbitals of the excited state (N=1)(N = 1), {ψa(1),ψb(1)}\{\psi_a^{(1)}, \psi_b^{(1)}\},

Mij(t)=⟨ψi(1)(t)∣P^0(t)∣ψj(1)(t)⟩,i,j∈{a,b}(30)M_{ij}(t) = \langle\psi_i^{(1)}(t) \lvert\hat{P}_0(t) \rvert\psi_j^{(1)}(t) \rangle,\quad i,j \in\{a,b\} \tag*{(30)}

The eigenvalues of M(t)\mathbf{M}(t) are the weights with which the corresponding orbital combinations project onto the occupied ground-state space, analogous to natural transition orbitals. [126, 127] The eigenvectors define linear combinations of the two active orbitals: one corresponding to an eigenvalue close to unity identifies a “hole” orbital lying almost entirely in the occupied ground-state space, and the other with an eigenvalue close to zero identifies a “particle” orbital nearly orthogonal to that space. We use the hole and particle DTOs to construct the excited-state singlet Φ1\Phi_1 according to Eq. (7). When the particle DTO has little weight in the occupied ground-state space, Φ1\Phi_1 has a small overlap with the ground state Φ0\Phi_0, although exact orthogonality is not imposed. The overlap matrix between these states at successive time steps (Eq. (31)) is then used to evaluate the time-derivative coupling matrix T\mathbf{T}, as described in § 3.3.

The projector P^0\hat{P}_0 depends only on the occupied subspace of Φ0\Phi_0, so the DTO basis is insensitive to unitary rotations among the occupied ground-state orbitals. Within the excited-state active space, the equal occupations fa=fb=1f_a = f_b = 1 ensure that the density ρ=∑nfn∣ψn∣2\rho= \sum_n f_n \lvert\psi_n \rvert^2 is invariant under unitary mixing of {a,b}\{a,b\}. The DTO transformation therefore leaves the density, energy and nuclear forces unchanged, while fixing the constructed many-electron wavefunctions used for overlaps. Explicit rotation is needed only in the active space, since a unitary rotation among doubly occupied orbitals changes a Slater determinant only by an overall phase.

After the particle–hole directions have been selected, each DTO still has an arbitrary phase. Independent diagonalization of Eq. (30) at each step can thus produce discontinuous phase changes and artificial spikes in the overlap-based couplings. We therefore align each DTO with the corresponding orbital at the previous step by choosing its phase so that their overlap is real and positive, provided the overlap is nonzero. This maintains a continuous gauge along the trajectory.

Numerical Evaluation of Nonadiabatic Couplings

The time-derivative coupling matrix T\mathbf{T} is needed at every step to propagate the spin-mapping variables according to Eq. (26). Since the many-electron wavefunctions constructed in the DTO gauge are expressed in terms of Slater determinants, their overlaps between successive time steps are directly available, as in overlap-based estimates of the time-derivative coupling. [78] We obtain T\mathbf{T} from these overlaps using local diabatization. [97, 109, 128, 129]

In local diabatization, the adiabatic basis at time tt is held fixed and used as a diabatic basis over a short interval, with the dynamics approximated within this two-state space. The change between the basis functions at tt and t+Δtt+\Delta t, including the phases carried by the propagated states, is measured by the overlap matrix

UMN=⟨ΦM(R(t))∣ΦN(R(t+Δt))⟩(31)U_{MN} = \langle\Phi_M(\mathbf{R}(t))|\Phi_N(\mathbf{R}(t+\Delta t))\rangle \tag*{(31)}

The overlap of two Slater determinants is the determinant of their occupied spin-orbital overlap matrix, which factorizes into one spatial-orbital determinant per spin channel. Applying this rule to Eqs. (6) and (7) (see Figure 1) gives [32, 77]

U=(Dgg22DghDgl2DhgDlgDhhDll+DhlDlh)(32)\mathbf{U} = \begin{pmatrix} \mathcal{D}_{gg}^{2} & \sqrt{2}\mathcal{D}_{gh}\mathcal{D}_{gl} \\ \sqrt{2}\mathcal{D}_{hg}\mathcal{D}_{lg} & \mathcal{D}_{hh}\mathcal{D}_{ll}+\mathcal{D}_{hl}\mathcal{D}_{lh} \end{pmatrix} \tag*{(32)}

Here gg denotes the occupied ground-state spatial orbitals; hh and ll contain the doubly occupied excited-state orbitals plus the hole and particle DTO, respectively. Each D\mathcal{D} is the determinant of a spatial-orbital overlap matrix, with the first subscript referring to tt and the second to t+Δtt + \Delta t. For example, Dgh=det⁡Ogh\mathcal{D}_{gh} = \det\mathbf{O}_{gh} with

Ogh=Cg†(t) S(R(t),R(t+Δt)) Ch(t+Δt)(33)\mathbf{O}_{gh} = \mathbf{C}_{g}^{\dagger}(t)\,\mathbf{S}\bigl(\mathbf{R}(t),\mathbf{R}(t+\Delta t)\bigr)\,\mathbf{C}_{h}(t+\Delta t) \tag*{(33)}

where Cg\mathbf{C}_{g} and Ch\mathbf{C}_{h} contain the corresponding AO coefficients, and Sμν(R,R′)=⟨ϕμ(R)∣ϕν(R′)⟩S_{\mu\nu}(\mathbf{R},\mathbf{R}') = \langle\phi_{\mu}(\mathbf{R})|\phi_{\nu}(\mathbf{R}')\rangle is the AO overlap matrix between the two geometries. The remaining orbital overlap matrices are constructed analogously.

As the nuclear geometry changes, the states at t+Δtt+\Delta t may acquire components outside the space spanned by the two states at tt, so their 2×22 \times2 overlap matrix need not be unitary. In the trajectories reported below, this deviation from unitarity appears as singular values 10−310^{-3}–10−410^{-4} below unity per step. Although small at each step, these deviations can accumulate over a trajectory. Using a non-unitary U\mathbf{U} in the matrix logarithm introduced below would give Tˉ\bar{\mathbf{T}} a Hermitian component, inconsistent with the anti-Hermitian couplings required for the norm-conserving two-state evolution underlying Eq. (26). We therefore project U\mathbf{U} onto the nearest unitary matrix through its singular-value decomposition, [130]

U=WΣZ†,Ucorr=WZ†(34)\mathbf{U} = \mathbf{W}\boldsymbol{\Sigma}\mathbf{Z}^{\dagger}, \qquad\mathbf{U}_{\mathrm{corr}} = \mathbf{W}\mathbf{Z}^{\dagger} \tag*{(34)}

where W\mathbf{W} and Z\mathbf{Z} are unitary and Σ\boldsymbol{\Sigma} holds the singular values.

Since Ucorr\mathbf{U}_{\mathrm{corr}} is unitary, its generator is the anti-Hermitian coupling matrix Tˉ\bar{\mathbf{T}}, provided the electronic states vary smoothly, such that Ucorr≈exp⁡(TˉΔt)\mathbf{U}_{\mathrm{corr}} \approx\exp(\bar{\mathbf{T}}\Delta t) over a short step, and we can obtain Tˉ\bar{\mathbf{T}} from [130]

\bar{\mathbf{T}} \approx\frac{\logm\left(\mathbf{U}_{\mathrm{corr}}\right)}{\Delta t} \tag*{(35)}

where \logm denotes the principal matrix logarithm. This gives the full coupling matrix, including its diagonal and off-diagonal elements in the complex basis.

Whereas Tˉ\bar{\mathbf{T}} is required at every step, the NACV d01\mathbf{d}_{01} is evaluated only at an attempted hop, to determine whether the hop is energetically allowed and, if so, to rescale the momentum.

Equation 31 uses the overlap logarithm to estimate how the electronic basis changes per unit time. The NACV describes the corresponding change per unit nuclear displacement. This correspondence motivates estimating the NACV from the spatial variation of the overlap logarithm, with the initial states and propagation interval held fixed. We evaluate it within the same local-diabaticization framework, in which Ucorr(R,R′;Δt)\mathrm{U}_{\mathrm{corr}}(\mathbf{R},\mathbf{R}';\Delta t) denotes the unitarized overlap between states at R\mathbf{R} and those propagated to R′\mathbf{R}' in a time Δt\Delta t by Eq. (9). The NACV is then estimated as

d01A(R)≈∇A′[log⁡m(Ucorr(R,R′;Δt))]01∣R′=R(36)\mathbf{d}^{A}_{01}(\mathbf{R}) \approx\nabla'_{A}\left[\log_{m}\left(\mathrm{U}_{\mathrm{corr}}(\mathbf{R},\mathbf{R}';\Delta t)\right)\right]_{01}\bigg|_{\mathbf{R}'=\mathbf{R}} \tag*{(36)}

Here ∇A′\nabla'_{A} denotes the gradient with respect to RA′\mathbf{R}'_{A}, with the initial states and Δt\Delta t held fixed. Numerically, this gradient is evaluated by central finite differences using geometries displaced by ±ΔR\pm\Delta R along each nuclear Cartesian coordinate, with the overlaps unitarized as in Eq. (34). The propagation interval Δt\Delta t used here is examined in § 4.4 and need not equal the trajectory time step.

According to the quantum adiabatic principle (§ 2.3), the electronic states need sufficient propagation time to follow their adiabatic states across the displacement. The extraction therefore requires both a sufficiently small ΔR\Delta R for the central difference and a sufficiently long propagation interval, or equivalently a large enough α\alpha by the reparametrization of Eq. (10).

The extracted d01\mathbf{d}_{01} is complex because the constructed many-electron wavefunctions carry real-time phases. However, we know that in principle there exists a basis in which the instantaneous adiabatic states may be chosen real, with an underlying real coupling dˉ01\bar{\mathbf{d}}_{01}. To first order in ΔR\Delta R and within the two-state space, propagation across the displacement multiplies all Cartesian components by the same complex factor z=∣z∣eiφz = |z|e^{i\varphi}, which depends only on αΔEΔt\alpha\Delta E\Delta t. Writing d01=zdˉ01\mathbf{d}_{01} = z\bar{\mathbf{d}}_{01}, the combination entering the rescaling force of Eq. (28) is then

Re⁡(d01)Sx+Im⁡(d01)Sy=∣z∣dˉ01(Sxcos⁡φ+Sysin⁡φ)(37)\operatorname{Re}(\mathbf{d}_{01})S_x+\operatorname{Im}(\mathbf{d}_{01})S_y=|z|\bar{\mathbf{d}}_{01}(S_x\cos\varphi+S_y\sin\varphi) \tag*{(37)}

For a nonzero scalar prefactor, this vector is parallel to dˉ01\bar{\mathbf{d}}_{01}, so zz changes its magnitude but not its axis within this approximation. No additional phase fixing is required. In the present implementation, we use Re⁡(d01)\operatorname{Re}(\mathbf{d}_{01}) alone, which is parallel to dˉ01\bar{\mathbf{d}}_{01} provided Re⁡z≠0\operatorname{Re}z\ne0. Note that the magnitude and sign of the NACV are irrelevant for the momentum rescaling and that only the direction is required.

Each such NACV evaluation for a system of NatomN_{\mathrm{atom}} atoms costs 6Natom6N_{\mathrm{atom}} additional short propagations—the most expensive step of the scheme—but it is only needed at attempted hops, which occur only a few times per trajectory. Nevertheless, analytical NACVs between the propagated states, along the lines available for other mean-field excited-state methods, [131, 132] would remove this finite-difference cost and are planned for future work.

Integration of the Coupled Dynamics

Algorithm 2 uses the propagated electronic states and the numerically evaluated couplings of § 3.3 to evolve the coupled two-state MASH dynamics. The nuclei are integrated with velocity Verlet, while the two electronic states are propagated independently at the midpoint geometry of each step and transported to its end using Algorithm 1. Their state-specific energies and time-derivative coupling matrix drive the spin evolution, while sign⁡(Sz)\operatorname{sign}(S_z) selects the active surface that supplies the nuclear force. Here FA=−∇AVa\mathbf{F}_{A}=-\nabla_{A}V_{a} denotes the smooth force on that surface, whereas the impulsive contribution in Eq. (28) is handled by momentum rescaling at a hop. The implementation uses the nonreversible local-diabatisation scheme of Ref. 109. A hop attempt is assessed at the end of the step, after both nuclear half kicks; the second half kick uses the force of the state active at the start of the step, and the active state and force passed to the next step are set only after the hop has been accepted or rejected. Updates for steps in which SzS_z does not change sign are symmetric under time reversal, but this symmetry is not imposed on steps containing a hop attempt. More efficient reversible schemes, which locate the crossing within the step and split the propagation there to apply momentum rescaling before continuing[109] are left for future work.

Algorithm 2 Coupled MASH dynamics over one time step Δt\Delta t, from R(t)\mathbf{R}(t) to R(t+Δt)\mathbf{R}(t+\Delta t).

input: positions R(t), momenta P(t), spin S⃗, states Φ₀, Φ₁, force F(t)

nuclei (Verlet)
    half kick: P_A ← P_A + (Δt/2) F_A
    drift:     R_A(t + Δt/2) ← R_A(t) + (Δt/2) P_A/M_A
               R_A(t + Δt)   ← R_A(t) + Δt P_A/M_A

electrons
    propagate Φ₀, Φ₁ to R(t + Δt) by Algorithm 1; obtain V₀, V₁ and forces

nuclei (Verlet)
    half kick: P_A ← P_A + (Δt/2) F_A, with F_A = −∇_A V_a(R(t + Δt)) and a the active
    state at t

spin
    T̄ ← log_m(U_corr)/Δt from DTO-gauge overlaps, Eqs. 28–31
    propagate S⃗ by Eq. 22; sign(S_z) selects the candidate state b

hop
    if b ≠ a then evaluate d₀₁ by Eq. 32
        if energetically allowed then rescale P along Re(d₀₁); a ← b
        else S⃗ ← −S⃗ (frustrated hop)
    F_A(t + Δt) ← −∇_A V_a(R(t + Δt))

return R(t + Δt), P(t + Δt), S⃗, Φ₀, Φ₁, F(t + Δt)

Results and Discussion

Computational Details

The method was implemented in an in-house Python package built on PySCF[88, 89] (version 2.12.1) and applied to trans-diimide (N2H2\mathrm{N_2H_2}) with the 6-31G* basis set. [133, 134, 135] Diimide is the simplest azo compound, and its excited states were characterized early by ab initio methods. [136] It provides a minimal model of the azo unit relevant to azobenzene photoisomerization. [137, 138] Two levels of theory are used throughout: real-time Hartree–Fock (RT-HF) and RT-TDDFT with the exchange-only local-density approximation (LDA). [139] We chose this functional as a demanding numerical test rather than for its accuracy. It gives small energy gaps in the strong-coupling region, where static SCF becomes unreliable and the dynamical propagation is most useful. RT-HF is the complementary case. Its static solutions converge reliably along most of the path and thus provide a reference against which the propagated energies are checked. In both cases, the electronic trajectories are initialized from converged spin-restricted SCF solutions for the ground state (N=0N = 0) and the first excited state (N=1N = 1), as described in Sec. 2.2. For comparison, static references are computed on exactly the same geometries as the propagated results, using the CIAH orbital optimizer for the ground state and DIIS with MOM for the excited state, with a convergence threshold of 10−1210^{-12} Ha and at most 500 iterations.

The dynamical propagation of the electronic degrees of freedom uses the Crank–Nicolson scheme of Eq. (19). Its predictor–corrector iteration is taken as converged when the largest change in the molecular-orbital coefficients between successive iterations falls below 10−510^{-5}, and it is accelerated by Anderson mixing. A larger α\alpha does in principle demand a smaller time step for a given accuracy, since the phase accumulated per step grows in proportion to α\alpha. In practice, the Anderson-accelerated iteration converges reliably up to the largest α\alpha used here, and the convergence of the ensemble results with respect to the time step is shown in the Supporting Information. For trajectory propagation, we use equal electronic and nuclear time steps of 0.4 a.u. (0.0097 fs) unless otherwise stated for specific validation tests. This step is small enough for the largest α\alpha used, and keeping it fixed makes trajectories computed with different α\alpha directly comparable. Such a small step is typical of conventional RT-TDDFT propagation[40, 41] and was chosen conservatively for the present proof-of-principle study. More efficient RT-TDDFT schemes with larger propagation intervals are available[140, 141] and could be adopted in future work.

Initial nuclear coordinates and velocities were generated by Wigner sampling[142, 143] at 300 K, using the harmonic approximation to the ground-state LDA potential-energy sur face (details in § 4.6). To examine the electronic propagation and the evaluation of the couplings in isolation, we first use prescribed nuclear paths. § 4.2 defines two paths of 800 a.u. (19.4 fs) through the strong-coupling region, one for RT-HF and one for RT-TDDFT; the RT-TDDFT path is also used in § 4.3 and § 4.5. Both paths start from the same snapshot, taken partway along a longer trajectory started from a Wigner-sampled initial condition on the ground state. Its geometry and nuclear velocities are provided in the Supporting Information to ensure the reproducibility of our results. Starting from the snapshot, each path is generated once by α=1\alpha= 1 Ehrenfest dynamics, in which the nuclei feel the mean-field force of the instantaneous excited-state density, [108, 144] and the midpoint and end-point geometries of every step are stored. Then, for a range of values of α\alpha, both electronic states are propagated along the stored trajectory without letting the nuclei respond. Any difference between two values of α\alpha thus comes from the electronic propagation alone. No hops are attempted, no momenta are rescaled, and the spin-mapping variables of § 4.5 evolve without acting back on the nuclei. The spin dynamics shown there is thus a diagnostic of the electronic evolution rather than a surface-hopping simulation.

For the full coupled dynamics in § 4.6, we use an ensemble of 100 trajectories drawn from the same Wigner distribution, each propagated from t=0t = 0 for 4000 a.u. (96.8 fs), which is long enough to observe the relaxation from the excited state to the ground state. Here the state-specific forces switch between the two surfaces according to the sign of SzS_z, as prescribed by Eq. (28), and hops and momentum rescaling are active.

Numerical Tests of Adiabatic-State Following

Figure 2 compares the dynamically propagated ground- and excited-state energies at several values of α\alpha with static SCF references computed along the same nuclear path. This comparison tests adiabatic-state following (§ 2.3). We first examine the RT-HF results in Figure 2(a). At α=0.1\alpha= 0.1, the electronic response is slow relative to the nuclear motion, thereby narrowing instead of widening the separation of electronic and nuclear frequencies; the propagated wavefunction thus mixes with higher-lying states and departs substantially from the static SCF reference. At α=1\alpha= 1, the propagated energies initially follow the adiabatic surfaces but depart from them on entering the strong-coupling region near t=10.4t = 10.4 fs. As α\alpha is increased further, the propagated energies converge systematically to the static solution. In particular, at α=40\alpha= 40 the timescales are well separated, and both propagated energies agree with the static references along the whole path to within 1% of the energy gap between the states.

Figure 2. Potential-energy profiles along trans-diimide nuclear paths obtained with (a) RT-HF and (b) RT-TDDFT. Dynamically propagated results at several values of α\alpha are compared with static SCF references. Solid and dashed curves denote the propagated ground- and excited-state energies, respectively; open circles and crosses denote the corresponding static references. Within each panel, all propagated curves and static references are computed on the same nuclear path, generated once by α=1\alpha= 1 Ehrenfest dynamics.

We next repeat the test with RT-TDDFT, a more demanding case because the relevant gaps are considerably smaller than at the RT-HF level. As Figure 2(b) shows, the ground- and excited-state surfaces come close to each other near t=9.9t = 9.9 fs. As in the RT-HF case, the propagated energies at α=0.1\alpha= 0.1 and 1 depart from the static references. At α≥10\alpha\ge10 the propagated energies converge towards the static references as α\alpha is increased, except in a region where the static ground-state solution leaves the lower branch. In this strong-coupling region standard DIIS failed to converge for every level shift [145] we tested; static solutions could be obtained only with the CIAH orbital optimizer, which avoids direct diagonalization of the Fock matrix.

Even where it converges, the static CIAH ground-state solution leaves the lower branch near t=8.9t = 8.9 fs and lies up to 0.65 eV above the propagated ground state until about t=10.6t = 10.6 fs, where it drops back abruptly, leaving an unphysical kink in the profile. The dynamical result at α=80\alpha= 80 instead follows a lower-energy, continuous branch through the same region. This suggests that the static solver converges to a higher-lying stationary solution, whereas the propagated wavefunction keeps continuity with the true ground state selected earlier in the trajectory. Since our aim in this work is the dynamical propagation, we have not attempted to fix the static solver to recover the lower branch, although more advanced static optimization schemes[146] could presumably locate it. The dynamic propagation, by contrast, follows it correctly without any adjustment of constraints or initial guesses. For the excited state, the energy deviation from the static reference is typically below 2 meV over the whole path for α≥10\alpha\ge10. For the ground state, the energy deviation after the strong-coupling region decreases from about 2.5 eV at α=1\alpha= 1 to at most about 50 meV at α≥40\alpha\ge40.

In this proof-of-principle demonstration of state following, we have not optimized the integration scheme for efficiency or assessed its performance systematically. Static Born–Oppenheimer dynamics can take much larger time steps but requires a fully converged SCF solution at each geometry, whereas each dynamical propagation step typically reaches self-consistency within a few predictor–corrector iterations, requiring only a small number of Fock-matrix constructions. Counting these constructions, which dominate the cost of both approaches, the total effort is of a similar order of magnitude for the system studied here. The main advantage demonstrated here is robustness rather than speed: the propagated trajectory follows the selected adiabatic state through a region where the static solvers either fail to converge or converge to different solutions. However, with an optimized implementation, we expect that the dynamical propagation scheme could outperform the static SCF in terms of efficiency, especially for large condensed-phase systems. [45, 46]

Orthogonality in the DTO Representation

Figure 3 compares two constructions of the ground- and excited-state many-electron wavefunctions from the same TDKS orbitals propagated with α=10\alpha= 10 along the RT-TDDFT/Ehrenfest path of Figure 2(b). When the wavefunctions are constructed directly from the propagated TDKS orbitals, their overlap stays below 0.04 far from the crossing, rises steeply on entering the strong-coupling region, peaks at 0.70, and fluctuates about 0.55 thereafter. The rise coincides with the strong-coupling region identified in § 4.2. However, in the DTO gauge, the overlap remains small: it never exceeds 1.6×10−41.6 \times10^{-4} and averages 2.0×10−52.0 \times10^{-5}, so the two wavefunctions remain nearly orthogonal even through the strong-coupling region.

The two eigenvalues of M(t)\mathbf{M}(t) remain within 10−410^{-4} of unity and zero throughout the trajectory. These values indicate that the active space consists of one orbital of hole character, relative to the occupied ground-state space, and one of particle character, nearly orthogonal to it, although the individual propagated orbitals can mix them as the nuclei move. The DTO particle–hole construction, [81] described in § 3.2, selects these combinations for the many-electron wavefunctions used in the overlap calculations. We repeated the comparison at α=1\alpha= 1, 10, 40 and 80. In each case, the DTO gauge yields similarly small overlaps while the TDKS overlap remains large. Note that near orthogonality alone is not sufficient to establish accurate adiabatic-state following: at α=1\alpha= 1, the DTO overlap remains small despite the substantial energy deviations seen in § 4.2.

Validation of the Nonadiabatic Coupling Vectors

We validate the finite-difference evaluation of the NACV d01\mathbf{d}_{01} with RT-HF benchmark calculations on trans-diimide. The procedure is identical for RT-TDDFT, but we chose RT-HF because its static solutions converge reliably at the displaced geometries. Both the dynamical and static vectors are evaluated by central differences with a displacement of ΔR=0.01\Delta R = 0.01 a.u. (5.3×10−35.3 \times10^{-3} Å) along each nuclear Cartesian degree of freedom, as in Eq. (36). In both calculations, the orbital overlaps are evaluated using the AO overlap matrix between the displaced geometries, as in Eq. (33). For the dynamical calculation, we fix α=1\alpha= 1 and vary the duration Δt\Delta t of the single Crank–Nicolson step across each displacement. The static reference is obtained from SCF solutions converged at the same displaced geometries; its direction is shown in Figure 4. We quantify the directional agreement by the normalized dot product between the real part of the dynamically computed vector, d01cal=Re⁡(d01)\mathbf{d}_{01}^{\mathrm{cal}} = \operatorname{Re}(\mathbf{d}_{01}), and the static reference d01ref\mathbf{d}_{01}^{\mathrm{ref}}. Table 1 shows that this dot product increases from 0.987 at Δt=0.05\Delta t = 0.05 a.u. to 0.999 at Δt=4\Delta t = 4 a.u., meaning that the direction is recovered almost perfectly.

Spatial direction of the trans-diimide nonadiabatic coupling vector

Figure 4. Spatial direction of the trans-diimide nonadiabatic coupling vector (NACV) evaluated with the local-diabatization scheme. The Cartesian coordinates are provided in the Supporting Information.

Δt\Delta t (a.u.)dref⋅dcal/(∣dref∣∣dcal∣)\mathbf{d}^{\mathrm{ref}} \cdot\mathbf{d}^{\mathrm{cal}} / (\lvert\mathbf{d}^{\mathrm{ref}} \rvert\lvert\mathbf{d}^{\mathrm{cal}} \rvert)
0.050.987
0.100.988
0.400.988
0.500.988
1.000.990
2.000.995
4.000.999

Table 1. Convergence of the computed NACV direction toward the static reference with increasing propagation interval Δt\Delta t, for RT-HF trans-diimide with a finite-difference displacement ΔR=0.01\Delta R = 0.01 a.u. (5.3×10−35.3 \times10^{-3} Å) and α=1\alpha= 1. Alignment is quantified by the normalized dot product, in which dcal\mathbf{d}^{\mathrm{cal}} is the real part Re⁡(d01)\operatorname{Re}(\mathbf{d}_{01}), so that both vectors are real and ∣d∣=(d⋅d)1/2\lvert\mathbf{d} \rvert= (\mathbf{d} \cdot\mathbf{d})^{1/2}.

This improvement in directional agreement with increasing Δt\Delta t is consistent with the quantum adiabatic principle of § 2.3. A single Crank–Nicolson step across a fixed displacement depends on α\alpha and Δt\Delta t only through their product αΔt\alpha\Delta t (Eq. (10)), so increasing Δt\Delta t at α=1\alpha= 1 is equivalent to increasing α\alpha at fixed Δt\Delta t. Increasing this product gives the electrons more time to respond to the displacement, bringing the propagated states closer to the adiabatic states at the displaced geometries. The propagation remains stable over the whole range examined. This comparison validates only the direction of d01\mathbf{d}_{01}. Its norm, ∣d01∣=(d01†d01)1/2\lvert\mathbf{d}_{01} \rvert= (\mathbf{d}_{01}^{\dagger}\mathbf{d}_{01})^{1/2}, oscillates rapidly with α\alpha and with the propagation interval in these tests.

Nonetheless, it is only the direction which is required by MASH.

Spin Dynamics on the Bloch Sphere

To test whether the time-derivative couplings extracted from the propagated states drive spin-vector dynamics that converge with α\alpha, we track S⃗(t)\vec{S}(t) on the Bloch sphere along the RT-TDDFT/Ehrenfest path of Figure 2(b). The spin vector starts at the north pole, (Sx,Sy,Sz)=(0,0,1)(S_x, S_y, S_z) = (0, 0, 1), which corresponds to the first excited state (N=1N = 1), and the computational settings are identical to those of Figure 2(b). As noted above, the nuclei follow the prescribed path, without hops or momentum rescaling. Note that the spin is propagated with the unscaled energies and couplings. Because the static CIAH ground-state solution leaves the propagated branch in the strong-coupling region (§ 4.2), overlaps between consecutive static solutions would not provide a consistent reference for the couplings there. We therefore do not include a static-reference curve in Figure 5.

Figure 5. Evolution of the spin-mapping variable SzS_z along the RT-TDDFT/Ehrenfest path of Figure 2(b) for several values of the scaling parameter α\alpha.

Figure 5 shows SzS_z along this path. For α≥10\alpha\ge10 the Bloch vector leaves the north pole near t=8.6t = 8.6 fs, as the molecule enters the strong-coupling region, and crosses the equator about 0.5 fs later. In particular, at α=1\alpha= 1 this transfer is delayed to 11.1 fs. However, the first-crossing times for the larger values of α\alpha are 9.01 fs for α=10\alpha= 10, 9.23 fs for α=40\alpha= 40 and 9.15 fs for α=80\alpha= 80, and are thus reasonably well converged.

After the nonadiabatic coupling region, larger values of α\alpha suppress the unphysical residual oscillations. With α=1\alpha= 1, SzS_z keeps oscillating with an amplitude close to unity, consistent with the energy deviations seen in § 4.2, where the propagated states depart from their adiabatic states at small α\alpha. In particular, over the last 4 fs of the path, the standard deviation of SzS_z falls from 0.25 at α=10\alpha= 10 to 0.006 at both α=40\alpha= 40 and 80. The averages over the same interval at α=40\alpha= 40 and 80 are −0.574-0.574 and −0.511-0.511, so the late-time SzS_z is nearly converged by α=40\alpha= 40. Residual oscillations at small α\alpha could generate additional hops in fewest-switches surface hopping, where the hopping probability is computed directly from the time-derivative coupling. In MASH a hop occurs only when SzS_z changes sign, so for hopping decisions these oscillations matter only if they carry SzS_z across the equator.

Nonadiabatic Dynamics

We now test the α\alpha dependence of the full coupled RT-TDDFT/MASH dynamics for simulating electronic relaxation and trans–cis isomerization. The nuclei now evolve under the force of the active electronic state instead of following the prescribed paths of § 4.2–§ 4.5, while both electronic states are propagated at each new nuclear geometry. The active surface switches between V0V_0 and V1V_1 according to the sign of SzS_z in Eq. (28), and accepted hops are accompanied by momentum rescaling. Here NACV evaluations use α=1\alpha= 1, ΔR=0.01\Delta R = 0.01 a.u., and a single Crank–Nicolson step of 4 a.u. (0.097 fs) across each displacement, while the trajectory uses its own α\alpha and a time step of 0.4 a.u.

The initial conditions were generated by the Wigner sampling described above, using harmonic frequencies from the ground-state LDA/6-31G* Hessian computed at its minimum-energy geometry. The resulting ensemble has an N=N bond length of 1.265 ±\pm 0.041 Å and an N–H bond length of 1.079 ±\pm 0.077 Å. These distributions are similar but not identical to those of the gas-phase Wigner ensemble of Mališ and Luber, which was obtained with PBE0[147] and for which the corresponding values are 1.229 ±\pm 0.036 and 1.037 ±\pm 0.069 Å.[82] Following the standard MASH prescription,[6] each trajectory was initialized in the first excited state, with the initial Bloch vector sampled from the upper hemisphere of the Bloch sphere with probability density proportional to ∣Sz∣|S_z|.

On-the-fly RT-TDDFT/MASH simulations were then performed over a range of scaling parameters. The time-dependent excited-state population is estimated from the ensemble of time-evolved spin-mapping variables as

P+(t)=⟨∣Sz(t=0)∣h(Sz(t))⟩⟨∣Sz(t=0)∣⟩=⟨h(Sz(t))⟩w(38)P_+(t) = \frac{\langle|S_z(t = 0)| h(S_z(t))\rangle}{\langle|S_z(t = 0)|\rangle} = \langle h(S_z(t))\rangle_{\mathrm{w}} \tag*{(38)}

where the latter incorporates the weighting factor ∣Sz(t=0)∣\lvert S_z(t = 0)\rvert directly into the sampling distribution.

Figure 6(a) shows that P+(t)P_+(t) converges once α≥10\alpha\ge10, with every ensemble started from the same set of initial conditions. At α=1\alpha= 1, the excited-state population does not decay completely, because the propagated ground state acquires considerable higher-state character, and its energy V0V_0 rises above V1V_1 for about 16% of the time on average, compared with less than 1% at α≥10\alpha\ge10. Averaging over the ensemble largely smooths out the large SzS_z oscillations of individual α=1\alpha= 1 trajectories (§ 4.5), which survive only as small fluctuations of the population about this residual value. By the end of the propagation, about 18% of the ensemble is still excited at α=1\alpha= 1, compared with at most 3% for α≥10\alpha\ge10.

At α=1\alpha= 1, the hopping histories identify frustrated relaxation and back-transfer as contributions to the residual excited-state population. Where V0>V1V_0 > V_1, transfer to the initially ground-state branch is uphill and is frustrated when the kinetic energy along the rescaling direction cannot cover the gap. The 18 trajectories remaining on the initially excited branch comprise eight with no accepted hops and ten repopulated by back-transfer. At α=80\alpha= 80, the ensemble averages about two accepted and eight frustrated hops per trajectory, with the frustrated-hop count dominated by a few trajectories with repeated attempts near Sz=0S_z = 0 (Supporting Information). Such repeated attempts are not inconsistent with converged ensemble population dynamics, as similar behavior has been reported in MASH simulations of a related system.[119]

We characterize the decay by fitting the averaged population to P+(t)≈P∞+(1−P∞)e−(t−t0)/τP_+(t) \approx P_\infty+ (1 - P_\infty)e^{-(t-t_0)/\tau} for t≥t0t \ge t_0, with P∞P_\infty allowing a nonzero long-time population. The fits use nonlinear least squares, with uncertainties estimated by bootstrap resampling[148] over the 100 trajectories; further details are given in the Supporting Information. The fitted parameters for all values of α\alpha are collected in Table 2. Once the simulations are converged for α≥10\alpha\ge10, t0t_0 and τ\tau agree within their statistical uncertainties. However, with α=1\alpha= 1, the decay starts about 2.5 fs later and proceeds faster, and the ensemble also differs significantly in its long-time limit.

α\alphat0t_0 (fs)τ\tau (fs)P∞P_{\infty}
111.6±0.711.6 \pm0.79.4±1.59.4 \pm1.50.169±0.0340.169 \pm0.034
109.1±0.79.1 \pm0.711.4±1.511.4 \pm1.50.017±0.0120.017 \pm0.012
408.8±0.78.8 \pm0.712.8±1.612.8 \pm1.60.017±0.0140.017 \pm0.014
809.4±0.69.4 \pm0.611.5±1.411.5 \pm1.40.018±0.0120.018 \pm0.012

Table 2. Fit of the averaged excited-state population to P+(t)≈P∞+(1−P∞)exp⁡[−(t−t0)/τ]P_{+}(t) \approx P_{\infty} + (1 - P_{\infty}) \exp[-(t - t_0)/\tau] for t≥t0t \ge t_0, for each value of α\alpha. Uncertainties are bootstrap standard errors from 5000 resamples over the 100 trajectories of the ensemble.

Mališ and Luber[82] quoted the lifetime as t0+τt_0 + \tau, for which we obtain 20.9±1.320.9 \pm1.3 fs at α=80\alpha= 80. This lifetime agrees well with other simulations of the same system, although none uses exactly the same level of theory. Mališ and Luber[82] report a PBE0-based biexponential decay for unshifted trans-diimide in the gas phase, with lifetimes of 37 and 152 fs. When they shift the Δ\DeltaSCF excited state down by 1.0 eV, so that the two surfaces come almost into contact at the twisted geometry, the fitted lifetime falls to 20 fs.[82] This value is relevant to our study, since the gaps in the present calculations are likewise small in the coupling region, and it is close to our result. Handt and co-workers,[149] using an RT-TDDFT/Ehrenfest method at the LDA/cc-pVDZ level, find the main nonadiabatic transition about 20 fs after excitation, which is also close to our result.

As in Eq. (38), we characterize isomerization by the instantaneous fraction of trajectories in the trans-like region, defined by ∣ϕHNNH∣>90∘|\phi_{\mathrm{HNNH}}| > 90^{\circ},

Ptrans(t)=⟨∣Sz(t=0)∣h(∣ϕHNNH(t)∣−90∘)⟩⟨∣Sz(t=0)∣⟩=⟨h(∣ϕHNNH(t)∣−90∘)⟩w(39)P_{\mathrm{trans}}(t) = \frac{\left\langle|S_z(t=0)|h\left(|\phi_{\mathrm{HNNH}}(t)|-90^{\circ}\right)\right\rangle}{\left\langle|S_z(t=0)|\right\rangle} = \left\langle h\left(|\phi_{\mathrm{HNNH}}(t)|-90^{\circ}\right)\right\rangle_{\mathrm{w}} \tag*{(39)}

where ϕHNNH\phi_{\mathrm{HNNH}} is the HNNH dihedral angle. Figure 6(b) shows the time-dependent trans-like fraction Ptrans(t)P_{\mathrm{trans}}(t). The four curves coincide for about the first 30 fs, so this structural observable is less sensitive to α\alpha at early times than the electronic population. The subsequent departure of the α=1\alpha= 1 ensemble is consistent with poorer state following (Figure 2). With α≥10\alpha\ge10, the curves agree within statistical fluctuations, and about 55% of the ensemble. occupies the cis-like region over the last 20 fs. Although the excited-state population has largely decayed by this time, both regions remain appreciably populated, consistent with a vibrationally excited ensemble.

Figure 6. Ensemble dynamics of trans-diimide from 100 RT-TDDFT/MASH trajectories. (a) Time-dependent excited-state population P+(t)P_{+}(t) of Eq. (38) at several values of the scaling parameter α\alpha. (b) Isomerization measure Ptrans(t)P_{\mathrm{trans}}(t) of Eq. (39).

Conclusions

In this work, we have introduced an on-the-fly framework that combines energy-scaled RT-TDDFT with MASH for first-principles nonadiabatic molecular dynamics without requiring a static electronic-structure calculation at each time step. Following the quantum adiabatic principle, we scale the electronic Hamiltonian by a factor α\alpha to widen the separation between electronic and nuclear timescales, so that the ground- and excited-state SCF branches can be followed independently by real-time propagation. The physical energies and forces are evaluated from the unscaled functional. To use these propagated states in surface hopping, we reformulate the two-state MASH equations in a complex adiabatic basis and use a DTO gauge to construct the many-electron wavefunctions entering the overlap calculations. Local diabatization and spatial finite differences then provide the time-derivative couplings and the NACV directions required for the coupled dynamics.

We tested the method on the photoinduced dynamics of trans-diimide. As expected, systematically increasing α\alpha improves agreement between the propagated energies and the static references where the latter are reliable. Unlike the static optimizer, which produces unphysical discontinuities, RT-TDDFT produces continuous energy surfaces through the nonadiabatic coupling region. The DTO construction yields nearly orthogonal wavefunctions, and the RT-HF benchmark validates the NACV direction needed for momentum rescaling. In the full MASH simulations, the electronic populations and isomerization behavior converge as α\alpha is increased. In particular, increasing α\alpha reduces the population remaining on the initially excited branch at long times, and the lifetime is consistent with earlier simulations. [82, 149]

This study is a proof of principle, with robustness as its main demonstrated advantage. Extension beyond two states is possible based on the multistate formulation of MASH. 122

The implementation has not been optimized for performance, and its computational efficiency relative to static SCF-based surface hopping and other nonadiabatic methods remains to be assessed. An adaptive scheme could use a large α\alpha only where the nonadiabaticity demands it, allowing larger time steps elsewhere. More efficient electronic propagation[140, 141] and a multiple-time-step integrator, with the electrons advanced in substeps,[2, 150, 151] would further reduce the cost. With these developments, our scheme could offer a more accurate surface-hopping alternative to RT-TDDFT/Ehrenfest simulations of large molecular and condensed-phase systems.[152, 153, 154, 155]

Supporting Information Available

Cartesian coordinates and nuclear velocities of the snapshot from which the two nuclear paths of § 4.2–§ 4.5 start; details of the fits to the excited-state population at each value of α\alpha; convergence of the excited-state population with respect to the integration time step; hop statistics of the ensembles.

Data and Software Availability

The implementation used in this work is released as an open-source Python package, TDMASH, available at https://github.com/ruiyiQM/TDMASH_pack. The repository contains the energy-scaled real-time propagators, the DTO gauge construction, the nonadiabatic-coupling and MASH modules described in § 2.5–§ 3.3, and input scripts that reproduce the single-trajectory and ensemble calculations reported here. The Wigner-sampled initial conditions of the 100-trajectory ensemble are distributed with the code, so that the results of § 4.6 can be regenerated from the same starting configurations.

Acknowledgement

R.Z. and J.O.R. thank Kasra Asnaashari and Shivank Nigam for testing some initial ideas related to this work. We further thank Yu-Chen Wang, Hannes Hoppe, Daniele Furlanetto, Jonathan R. Mannouch, Johan E. Runeson and George Trenins for helpful discussions of the MASH algorithm, and Chao Lian for advice on the implementation of RT-TDDFT.

TOC Graphic

Figure 7 (PDF p. 59).

References

  1. [1]González, L., Lindh, R., Eds. Quantum Chemistry and Dynamics of Excited States: Methods and Applications; Wiley, 2021.DOI
  2. [2]Li, X.; Tully, J. C.; Schlegel, H. B.; Frisch, M. J. Ab initio Ehrenfest dynamics. The Journal of chemical physics 2005, 123, 084106.
  3. [3]Tully, J. C. Perspective: Nonadiabatic dynamics theory. J. Chem. Phys. 2012, 137, 22A301.
  4. [4]Parandekar, P. V.; Tully, J. C. Detailed balance in Ehrenfest mixed quantum-classical dynamics. J. Chem. Theory Comput. 2006, 2, 229–235.DOI
  5. [5]Tully, J. C. Molecular dynamics with electronic transitions. The Journal of Chemical Physics 1990, 93, 1061–1071.
  6. [6]Mannouch, J. R.; Richardson, J. O. A mapping approach to surface hopping. J. Chem. Phys. 2023, 158, 104111.
  7. [7]Richardson, J. O.; Lawrence, J. E.; Mannouch, J. R. Nonadiabatic Dynamics with the Mapping Approach to Surface Hopping (MASH). Annu. Rev. Phys. Chem. 2025, 76, 663–687.
  8. [8]Mai, S.; Marquetand, P.; González, L. Nonadiabatic dynamics: The SHARC approach. WIREs Comput. Mol. Sci. 2018, 8, e1370.
  9. [9]Crespo-Otero, R.; Barbatti, M. Recent advances and perspectives on nonadiabatic mixed quantum–classical dynamics. Chemical Reviews 2018, 118, 7026–7068.DOI
  10. [12]Hohenberg, P.; Kohn, W. Inhomogeneous electron gas. Physical review 1964, 136, B864.
  11. [13]Parr, R. G.; Yang, W. Density-functional theory of the electronic structure of molecules. Annual review of physical chemistry 1995, 46, 701–728.
  12. [14]Kohn, W.; Sham, L. J. Self-consistent equations including exchange and correlation effects. Physical review 1965, 140, A1133.
  13. [15]Casida, M. E. Recent Advances In Density Functional Methods: (Part I); World Scientific, 1995; pp 155–192.
  14. [16]Ullrich, C. A. Time-Dependent Density-Functional Theory: Concepts and Applications; Oxford University Press: Oxford, 2012.
  15. [17]Tapavicza, E.; Tavernelli, I.; Rothlisberger, U. Trajectory surface hopping within linear response time-dependent density-functional theory. Physical review letters 2007, 98, 023001.
  16. [18]Levine, B. G.; Ko, C.; Quenneville, J.; Martínez, T. J. Conical intersections and double excitations in time-dependent density functional theory. Molecular Physics 2006, 104, 1039–1051.
  17. [19]Yang, W.; Ayers, P. W. Foundation for the ΔSCF Approach in Density Functional Theory. arXiv preprint arXiv:2403.04604 2024,arxiv.org/abs/2403.04604
  18. [20]Hait, D.; Head-Gordon, M. Excited state orbital optimization via minimizing the square of the gradient: General approach and application to singly and doubly excited states via density functional theory. Journal of chemical theory and computation 2020, 16, 1699–1710.DOI
  19. [21]Schmerwitz, Y. L.; Ivanov, A. V.; Jónsson, E. Ö.; Jónsson, H.; Levi, G. Variational density functional calculations of excited states: Conical intersection and avoided crossing in ethylene bond twisting. The Journal of Physical Chemistry Letters 2022, 13, 3990–3999.DOI
  20. [22]Vigneshwaran, V.; Basumatary, S.; Beypi, C.; Akshay, C.; Ghosh, S. Why Variational Density Functional Theory Is More Accurate Than Time-Dependent Density Functional Theory for Certain” Difficult” Excited States? Journal of chemical theory and computation 2026, Pradhan, E.; Sato, K.; Akimov, A. V. Non-adiabatic molecular dynamics with ΔSCF excited states. Journal of Physics: Condensed Matter 2018, 30, 484002.
  21. [24]Hanasaki, K.; de Jong, T. F.; Komarov, K.; Kumar, R.; Mališ, M.; Mattiat, J.; Hernandez-Segura, L. I.; Schreder, L.; Sinyavskiy, A.; Luber, S. Exploring Excited-State Electronic Structure, Spectroscopy, and Nonadiabatic Dynamics with CP2K’s Multifaceted Approach. The Journal of Physical Chemistry A 2025, 129, 7313–7344.DOI
  22. [25]Vandaele, E.; Mališ, M.; Luber, S. The ΔSCF method for non-adiabatic dynamics of systems in the liquid phase. The Journal of Chemical Physics 2022, 156, 130901.DOI
  23. [26]Allen, W. D.; Schaefer III, H. F. An examination of the 2 1 A 1 states of formaldehyde and ketene including analytic configuration interaction energy first derivatives for singlet excited electronic states of the same symmetry as the ground state. The Journal of chemical physics 1987, 87, 7076–7095.DOI
  24. [27]Gilbert, A. T.; Besley, N. A.; Gill, P. M. Self-consistent field calculations of excited states using the maximum overlap method (MOM). The Journal of Physical Chemistry A 2008, 112, 13164–13171.DOI
  25. [28]Carter-Fenk, K.; Herbert, J. M. State-targeted energy projection: A simple and robust approach to orbital relaxation of non-Aufbau self-consistent field solutions. Journal of Chemical Theory and Computation 2020, 16, 5067–5082.
  26. [29]Kumar, C.; Luber, S. Robust ΔSCF calculations with direct energy functional minimization methods and STEP for molecules and materials. The Journal of Chemical Physics 2022, 156.DOI
  27. [30]Levi, G.; Ivanov, A. V.; Jónsson, H. Variational density functional calculations of excited states via direct optimization. Journal of Chemical Theory and Computation 2020, 16, 6968–6982.
  28. [31]Frank, I.; Hutter, J.; Marx, D.; Parrinello, M. Molecular dynamics in low-spin excited states. The Journal of chemical physics 1998, 108, 4060–4069.
  29. [32]Hait, D.; Zhu, T.; McMahon, D. P.; Van Voorhis, T. Prediction of excited-state energies and singlet–triplet gaps of charge-transfer states using a restricted open-shell Kohn–Sham approach. Journal of chemical theory and computation 2016, 12, 3353–3359.DOI
  30. [33]Ye, H.-Z.; Welborn, M.; Ricke, N. D.; Van Voorhis, T. σ-SCF: A direct energy-targeting method to mean-field excited states. The Journal of chemical physics 2017, 147.DOI
  31. [34]Ziegler, T.; Seth, M.; Krykunov, M.; Autschbach, J.; Wang, F. On the relation between time-dependent and variational density functional theory approaches for the determination of excitation energies and transition moments. The Journal of Chemical Physics 2009, 130.DOI
  32. [35]Evangelista, F. A.; Shushkov, P.; Tully, J. C. Orthogonality constrained density functional theory for electronic excited states. The Journal of Physical Chemistry A 2013, 117, 7378–7392.
  33. [36]Ramos, P.; Pavanello, M. Low-lying excited states by constrained DFT. The Journal of Chemical Physics 2018, 148.DOI
  34. [37]Runge, E.; Gross, E. K. Density-functional theory for time-dependent systems. Physical review letters 1984, 52, 997.
  35. [38]Marques, M. A. A.; Maitra, N. T.; Nogueira, F. M.; Gross, E. K.; Rubio, A. Fundamentals of time-dependent density functional theory; Springer, 2012; Vol. 837.
  36. [39]Yabana, K.; Bertsch, G. F. Time-dependent local-density approximation in real time. Physical Review B 1996, 54, 4484.
  37. [40]Li, X.; Govind, N.; Isborn, C.; DePrince III, A. E.; Lopata, K. Real-time time-dependent electronic structure theory. Chemical Reviews 2020, 120, 9951–9993.
  38. [41]Xu, J.; Carney, T. E.; Zhou, R.; Shepard, C.; Kanai, Y. Real-time time-dependent density functional theory for simulating nonequilibrium electron dynamics. Journal of the American Chemical Society 2024, 146, 5011–5029.
  39. [42]Ullrich, C. A. A snapshot of time-dependent density-functional theory. APL Computational Physics 2025, 1.
  40. [43]Zhou, R.; Yost, D. C.; Kanai, Y. First-principles demonstration of nonadiabatic Thouless pumping of electrons in a molecular system. The Journal of Physical Chemistry Letters 2021, 12, 4496–4503.DOI
  41. [44]Provorse, M. R.; Isborn, C. M. Electron dynamics with real-time time-dependent density functional theory. International Journal of Quantum Chemistry 2016, 116, 739–749.
  42. [45]Alonso, J. L.; Andrade, X.; Echenique, P.; Falceto, F.; Prada-Gracia, D.; Rubio, A. Efficient formalism for large-scale ab initio molecular dynamics based on time-dependent density functional theory. Physical review letters 2008, 101, 096403.
  43. [46]Andrade, X.; Castro, A.; Zueco, D.; Alonso, J.; Echenique, P.; Falceto, F.; Rubio, A. Modified Ehrenfest formalism for efficient large-scale ab initio molecular dynamics. Journal of chemical theory and computation 2009, 5, 728–742.DOI
  44. [47]Castro, A.; Marques, M. A.; Alonso, J. A.; Bertsch, G. F.; Rubio, A. Excited states dynamics in time-dependent density functional theory: high-field molecular dissociation and harmonic generation. The European Physical Journal D-Atomic, Molecular, Optical and Plasma Physics 2004, 28, 211–218.DOI
  45. [48]Meng, S.; Kaxiras, E. Real-time, local basis-set implementation of time-dependent density functional theory for excited state dynamics simulations. The Journal of chemical physics 2008, 129.DOI
  46. [49]Wang, F.; Yam, C. Y.; Hu, L.; Chen, G. Time-dependent density functional theory based Ehrenfest dynamics. The Journal of chemical physics 2011, 135.
  47. [50]Alonso, J. L.; Castro, A.; Echenique, P.; Rubio, A. In Fundamentals of Time-Dependent Density Functional Theory; Marques, M. A. L., Maitra, N. T., Nogueira, F. M. S., Gross, E. K. U., Rubio, A., Eds.; Springer: Heidelberg, 2012.
  48. [51]Ojanperä, A.; Havu, V.; Lehtovaara, L.; Puska, M. Nonadiabatic Ehrenfest molecular dynamics within the projector augmented-wave method. J. Chem. Phys. 2012, 136, 144103.DOI
  49. [52]Kolesov, G.; Granas, O.; Hoyt, R.; Vinichenko, D.; Kaxiras, E. Real-time TD-DFT with classical ion dynamics: Methodology and applications. Journal of Chemical Theory and Computation 2016, 12, 466–476.DOI
  50. [53]Bonafé, F. P.; Aradi, B.; Hourahine, B.; Medrano, C. R.; Hernández, F. J.; Frauenheim, T.; Sánchez, C. G. A real-time time-dependent density functional tight-binding implementation for semiclassical excited state electron–nuclear dynamics and pump–probe spectroscopy simulations. Journal of Chemical Theory and Computation 2020, 16, 4454–4469.
  51. [54]Shepard, C.; Zhou, R.; Yost, D. C.; Yao, Y.; Kanai, Y. Simulating electronic excitation and dynamics with real-time propagation approach to TDDFT within plane-wave pseudopotential formulation. The Journal of Chemical Physics 2021, 155, 100901.DOI
  52. [55]Xu, J.; Zhou, R.; Blum, V.; Li, T. E.; Hammes-Schiffer, S.; Kanai, Y. First-Principles Approach for Coupled Quantum Dynamics of Electrons and Protons in Heterogeneous Systems. Phys. Rev. Lett. 2023, 131, 238002.
  53. [56]Wang, L.-W. Natural orbital branching scheme for time-dependent density functional theory nonadiabatic simulations. The Journal of Physical Chemistry A 2020, 124, 9075–9087.DOI
  54. [57]Kang, J.; Wang, L.-W. Nonadiabatic molecular dynamics with decoherence and detailed balance under a density matrix ensemble formalism. Physical Review B 2019, 99, 224303.DOI
  55. [58]Abedi, A.; Maitra, N. T.; Gross, E. K. Exact factorization of the time-dependent electron-nuclear wave function. Physical Review Letters 2010, 105, 123002.
  56. [59]Ha, J.-K.; Min, S. K. Independent trajectory mixed quantum-classical approaches based on the exact factorization. The Journal of chemical physics 2022, 156.DOI
  57. [60]Han, D.; Ha, J.-K.; Min, S. K. Real-space and real-time propagation for correlated electron–nuclear dynamics based on exact factorization. Journal of Chemical Theory and Computation 2023, 19, 2186–2197.DOI
  58. [61]Craig, C. F.; Duncan, W. R.; Prezhdo, O. V. Trajectory Surface Hopping in the Time-Dependent Kohn-Sham Approach? format?; for Electron-Nuclear Dynamics. Physical review letters 2005, 95, 163001.
  59. [62]Fischer, S. A.; Habenicht, B. F.; Madrid, A. B.; Duncan, W. R.; Prezhdo, O. V. Regarding the validity of the time-dependent Kohn–Sham approach for electron-nuclear dynamics via trajectory surface hopping. The Journal of Chemical Physics 2011, 134.DOI
  60. [63]Shenvi, N.; Roy, S.; Tully, J. C. Nonadiabatic dynamics at metal surfaces: Independent-electron surface hopping. The Journal of chemical physics 2009, 130.DOI
  61. [64]Akimov, A. V.; Prezhdo, O. V. The PYXAID program for non-adiabatic molecular dynamics in condensed matter systems. Journal of chemical theory and computation 2013, 9, 4959–4972.DOI
  62. [65]Zheng, Q.; Chu, W.; Zhao, C.; Zhang, L.; Guo, H.; Wang, Y.; Jiang, X.; Zhao, J. Ab initio nonadiabatic molecular dynamics investigations on the excited carriers in condensed matter systems. Wiley Interdisciplinary Reviews: Computational Molecular Science 2019, 9, e1411.DOI
  63. [66]Liu, X.-Y.; Chen, W.-K.; Fang, W.-H.; Cui, G. Nonadiabatic dynamics simulations for photoinduced processes in molecules and semiconductors: Methodologies and applications. Journal of Chemical Theory and Computation 2023, 19, 8491–8522.DOI
  64. [67]Wang, L.; Akimov, A.; Prezhdo, O. V. Recent progress in surface hopping: 2011–2015. The journal of physical chemistry letters 2016, 7, 2100–2112.DOI
  65. [68]Car, R.; Parrinello, M. Unified approach for molecular dynamics and density-functional theory. Physical review letters 1985, 55, 2471.
  66. [69]Marx, D.; Hutter, J. Ab Initio Molecular Dynamics: Basic Theory and Advanced Methods; Cambridge University Press: Cambridge, 2009.
  67. [70]Hutter, J. Car–Parrinello molecular dynamics. Wiley Interdisciplinary Reviews: Computational Molecular Science 2012, 2, 604–612.
  68. [71]Kühne, T. D. Second generation Car–Parrinello molecular dynamics. Wiley Interdisciplinary Reviews: Computational Molecular Science 2014, 4, 391–406.
  69. [72]Niklasson, A. M. Extended born-oppenheimer molecular dynamics. Physical review letters 2008, 100, 123004.
  70. [73]Pastore, G.; Smargiassi, E.; Buda, F. Theory of ab initio molecular-dynamics calculations. Physical Review A 1991, 44, 6334.
  71. [74]Bjorgaard, J. A.; Sheppard, D.; Tretiak, S.; Niklasson, A. Extended Lagrangian excited state molecular dynamics. Journal of Chemical Theory and Computation 2018, 14, 799–806.
  72. [75]Fajen, O. J.; Gränås, O.; Martínez, T. J.; Niklasson, A. Shadow excited state molecular dynamics with the\DeltaSCF method. arXiv preprint arXiv:2507.21461 2025,
  73. [76]Tawfik, S. A.; Walsh, T. R. The hole mass in Car–Parrinello molecular dynamics: insights into the dynamics of excitation. Chemical Science 2025, 16, 10454–10464.DOI
  74. [77]Doltsinis, N. L.; Marx, D. Nonadiabatic car-parrinello molecular dynamics. Physical review letters 2002, 88, 166402.
  75. [78]Hammes-Schiffer, S.; Tully, J. C. Proton transfer in solution: Molecular dynamics with quantum transitions. The Journal of chemical physics 1994, 101, 4657–4667.
  76. [79]Kowalczyk, T.; Tsuchimochi, T.; Chen, P.-T.; Top, L.; Van Voorhis, T. Excitation energies and Stokes shifts from a restricted open-shell Kohn–Sham approach. The Journal of chemical physics 2013, 138.DOI
  77. [80]Büchel, R.; Álvarez, L.; Grage, J.; Maniscalco, D.; Frank, I. On the Simulation of Photoreactions Using Restricted Open-Shell Kohn–Sham Theory. Molecules 2024, 29, 4509.DOI
  78. [81]Zhou, R.; Kanai, Y. Dynamical transition orbitals: A particle–hole description in real-time TDDFT dynamics. The Journal of Chemical Physics 2021, 154.DOI
  79. [82]Malis, M.; Luber, S. Trajectory surface hopping nonadiabatic molecular dynamics with Kohn–Sham ΔSCF for condensed-phase systems. J. Chem. Theory Comput. 2020, 16, 4071–4086.DOI
  80. [83]Parr, R. G. Density functional theory of atoms and molecules. Horizons of quantum chemistry: proceedings of the third international congress of quantum chemistry held at Kyoto, Japan, October 29-November 3, 1979. 1989; pp 5–15.
  81. [84]Pulay, P. Convergence acceleration of iterative sequences. The case of SCF iteration. Chemical physics letters 1980, 73, 393–398.DOI
  82. [85]Head-Gordon, M.; Pople, J. A. Optimization of wave function and geometry in the finite basis Hartree-Fock method. The Journal of Physical Chemistry 1988, 92, 3063–3069.DOI
  83. [86]Van Voorhis, T.; Head-Gordon, M. A geometric approach to direct minimization. Molecular Physics 2002, 100, 1713–1721.
  84. [87]Sun, Q. Co-iterative augmented Hessian method for orbital optimization. arXiv preprint arXiv:1610.08423 2016,arxiv.org/abs/1610.08423
  85. [88]Sun, Q.; Berkelbach, T. C.; Blunt, N. S.; Booth, G. H.; Guo, S.; Li, Z.; Liu, J.; McClain, J. D.; Sayfutyarova, E. R.; Sharma, S., et al. PySCF: the Python-based simulations of chemistry framework. Wiley Interdisciplinary Reviews: Computational Molecular Science 2018, 8, e1340.arxiv.org/abs/1701.08223
  86. [89]Sun, Q.; Zhang, X.; Banerjee, S.; Bao, P.; Barbry, M.; Blunt, N. S.; Bogdanov, N. A.; Booth, G. H.; Chen, J.; Cui, Z.-H., et al. Recent developments in the PySCF program package. The Journal of chemical physics 2020, 153.DOI
  87. [90]Ziegler, T.; Rauk, A.; Baerends, E. J. On the calculation of multiplet energies by the Hartree-Fock-Slater method. Theoretica chimica acta 1977, 43, 261–271.
  88. [91]Born, M.; Fock, V. Beweis des adiab atensatzes. Zeitschrift für Physik 1928, 51, 165–180.
  89. [92]Kato, T. On the adiabatic theorem of quantum mechanics. Journal of the Physical Society of Japan 1950, 5, 435–439.
  90. [93]Griffiths, D. J.; Schroeter, D. F. Introduction to quantum mechanics; Cambridge university press, 2018.
  91. [94]Landau, L. D. Zur Theorie der Energieübertragung. II. Phys. Z. Sowjetunion 1932, 2, 46.
  92. [95]Zener, C. Non-adiabatic crossing of energy levels. Proc. R. Soc. Lond. A 1932, 137, 696–702.
  93. [96]Miyamoto, Y.; Tateyama, Y.; Oyama, N.; Ohno, T. Conservation of the pure adiabatic state in Ehrenfest dynamics of the photoisomerization of molecules. Scientific reports 2015, 5, 18220.DOI
  94. [97]Granucci, G.; Persico, M.; Toniolo, A. Direct semiclassical simulation of photochemical processes with semiempirical wave functions. The Journal of Chemical Physics 2001, 114, 10608–10615.DOI
  95. [98]Wang, L.; Prezhdo, O. V. A simple solution to the trivial crossing problem in surface hopping. The journal of physical chemistry letters 2014, 5, 713–719.DOI
  96. [99]Mannouch, J. R. Ensuring Complete Consistency between Hops and Wavefunction Propagation in Fewest-Switches Surface Hopping. arXiv preprint arXiv:2610.18743 2026,arxiv.org/abs/2610.18743
  97. [100]Artacho, E.; O’Regan, D. D. Quantum mechanics in an evolving Hilbert space. Physical Review B 2017, 95, 115155.
  98. [101]Tomfohr, J. K.; Sankey, O. F. Time-Dependent Simulation of Conduction through a Molecule. Physica Status Solidi (b) 2001, 226, 115–123.
  99. [102]Crank, J.; Nicolson, P. A practical method for numerical evaluation of solutions of partial differential equations of the heat-conduction type. Mathematical proceedings of the Cambridge philosophical society. 1947; pp 50–67.
  100. [103]Castro, A.; Marques, M. A.; Rubio, A. Propagators for the time-dependent Kohn–Sham equations. The Journal of Chemical Physics 2004, 121, 3425–3433.
  101. [104]Zhu, Y.; Herbert, J. M. Self-consistent predictor/corrector algorithms for stable and efficient integration of the time-dependent Kohn-Sham equation. J. Chem. Phys. 2018, 148, 044117.DOI
  102. [105]Anderson, D. G. Iterative procedures for nonlinear integral equations. Journal of the ACM (JACM) 1965, 12, 547–560.
  103. [106]Feynman, R. P. Forces in molecules. Physical review 1939, 56, 340.
  104. [107]Pulay, P. Ab initio calculation of force constants and equilibrium geometries in polyatomic molecules: I. Theory. Molecular Physics 1969, 17, 197–204.DOI
  105. [108]Xu, J.; Zhou, R.; Li, T. E.; Hammes-Schiffer, S.; Kanai, Y. Lagrangian formulation of nuclear–electronic orbital Ehrenfest dynamics with real-time TDDFT for extended periodic systems. The Journal of Chemical Physics 2024, 161.
  106. [109]Geuther, J. A.; Asnaashari, K.; Richardson, J. O. Time-Reversible Implementation of MASH for Efficient Nonadiabatic Molecular Dynamics. J. Chem. Theory Comput. 2025, 21, 2179–2188.
  107. [110]Granucci, G.; Persico, M. Critical appraisal of the fewest switches algorithm for surface hopping. J. Chem. Phys. 2007, 126, 134114.
  108. [111]Subotnik, J. E.; Jain, A.; Landry, B.; Petit, A.; Ouyang, W.; Bellonzi, N. Understanding the surface hopping view of electronic transitions and decoherence. Annu. Rev. Phys. Chem. 2016, 67, 387–417.
  109. [112]Amati, G.; Mannouch, J. R.; Richardson, J. O. Detailed balance in mixed quantum–classical mapping approaches. J. Chem. Phys. 2023, 159, 214114.
  110. [113]Lawrence, J. E.; Mannouch, J. R.; Richardson, J. O. Recovering Marcus Theory Rates and Beyond without the Need for Decoherence Corrections: The Mapping Approach to Surface Hopping. J. Phys. Chem. Lett. 2024, 15, 707–716.
  111. [114]Ghamari, D.; Richardson, J. O. Nonadiabatic rare events from transition-path sampling of MASH trajectories. The Journal of Chemical Physics 2026, 164, 174101.
  112. [115]Yang, X.; Reiner, M. M.; Bachmair, B.; González, L.; Dietschreit, J. C. B.; Dellago, C. NATPS: Nonadiabatic Transition Path Sampling Using the Time-Reversible Mapping Approach to Surface Hopping. The Journal of Physical Chemistry Letters 2026, 17, 6014–6023.DOI
  113. [116]Mannouch, J. R.; Kelly, A. Toward a Correct Description of Initial Electronic Coherence in Nonadiabatic Dynamics Simulations. J. Phys. Chem. Lett. 2024, 15, 11687–11695.DOI
  114. [117]Furlanetto, D.; Richardson, J. O. Simulating electronic coherences induced by conical intersections using MASH: Application to attosecond X-ray spectroscopy. J. Phys. Chem. Lett. 2025, 16, 6794–6800.
  115. [118]Asnaashari, K.; Richardson, J. O. Open quantum–classical systems: A hybrid MASH master equation. The Journal of Chemical Physics 2026, 164.
  116. [119]Mannouch, J. R.; Kelly, A. Quantum quality with classical cost: Ab initio nonadiabatic dynamics simulations using the mapping approach to surface hopping. The Journal of Physical Chemistry Letters 2024, 15, 5814–5823.
  117. [120]Lawrence, J. E.; Ansari, I. M.; Mannouch, J. R.; Manae, M. A.; Asnaashari, K.; Kelly, A.; Richardson, J. O. A MASH simulation of the photoexcited dynamics of cyclobutanone. J. Chem. Phys. 2024, 160, 174306.
  118. [121]Hutton, L.; Moreno Carrascosa, A.; Prentice, A. W.; Simmermacher, M.; Runeson, J. E.; Paterson, M. J.; Kirrander, A. Using a multistate mapping approach to surface hopping to predict the ultrafast electron diffraction signal of gas-phase cyclobutanone. The Journal of Chemical Physics 2024, 160. Lawrence, J. E.; Mannouch, J. R.; Richardson, J. O. A Size-Consistent Multi-State Mapping Approach to Surface Hopping. J. Chem. Phys. 2024, 160, 244112.
  119. [123]Runeson, J. E.; Manolopoulos, D. E. A multi-state mapping approach to surface hopping. J. Chem. Phys. 2023, 159, 094115.
  120. [124]Botticelli, S.; Accomasso, D.; Granucci, G. A semi-focused multi-state variant of the mapping approach to surface hopping. The Journal of Chemical Physics 2025, 163, 224124.DOI
  121. [125]Furlanetto, D.; Richardson, J. O. To be submitted. in preparation 2026,
  122. [126]Li, Y.; Ullrich, C. Time-dependent transition density matrix. Chemical Physics 2011, 391, 157–163.
  123. [127]Martin, R. L. Natural transition orbitals. Journal of Chemical Physics 2003, 118, 4775–4777.
  124. [128]Plasser, F.; Granucci, G.; Pittner, J.; Barbatti, M.; Persico, M.; Lischka, H. Surface hopping dynamics using a locally diabatic formalism: Charge transfer in the ethylene dimer cation and excited state dynamics in the 2-pyridone dimer. The Journal of chemical physics 2012, 137.DOI
  125. [129]Meek, G. A.; Levine, B. G. Evaluation of the time-derivative coupling for accurate electronic state transition probabilities from numerical simulations. The journal of physical chemistry letters 2014, 5, 2351–2356.
  126. [130]Jain, A.; Alguire, E.; Subotnik, J. E. An efficient, augmented surface hopping algorithm that includes decoherence for use in large-scale simulations. Journal of chemical theory and computation 2016, 12, 5256–5268.
  127. [131]Billeter, S. R.; Egli, D. Calculation of nonadiabatic couplings with restricted open-shell Kohn-Sham density-functional theory. The Journal of chemical physics 2006, 125.
  128. [132]Ramos, P.; Pavanello, M. Nonadiabatic couplings from a variational excited state method based on constrained DFT. The Journal of Chemical Physics 2021, 154.DOI
  129. [133]Ditchfield, R.; Hehre, W. J.; Pople, J. A. Self-consistent molecular-orbital methods. IX. An extended Gaussian-type basis for molecular-orbital studies of organic molecules. The Journal of Chemical Physics 1971, 54, 724–728.DOI
  130. [134]Hehre, W. J.; Ditchfield, R.; Pople, J. A. Self-consistent molecular orbital methods. XII. Further extensions of Gaussian-type basis sets for use in molecular orbital studies of organic molecules. The Journal of Chemical Physics 1972, 56, 2257–2261.DOI
  131. [135]Hariharan, P. C.; Pople, J. A. The influence of polarization functions on molecular orbital hydrogenation energies. Theoretica chimica acta 1973, 28, 213–222.DOI
  132. [136]Vasudevan, K.; Peyerimhoff, S. D.; Buenker, R. J.; Kammer, W. E.; Hsu, H.-I. Theoretical study of the electronic spectrum of diimide by ab initio methods. Chemical Physics 1975, 7, 187–209.DOI
  133. [137]Bandara, H. D.; Burdette, S. C. Photoisomerization in different classes of azobenzene. Chemical Society Reviews 2012, 41, 1809–1825.
  134. [138]Ciminelli, C.; Granucci, G.; Persico, M. The photoisomerization mechanism of azobenzene: A semiclassical simulation of nonadiabatic dynamics. Chemistry–A European Journal 2004, 10, 2327–2341.
  135. [139]Slater, J. C. The Self-Consistent Field for Molecules and Solids: Quantum Theory of Molecules and Solids, Vol. 4; McGraw-Hill: New York, 1974.
  136. [140]Jia, W.; An, D.; Wang, L.-W.; Lin, L. Fast real-time time-dependent density functional theory calculations with the parallel transport gauge. Journal of chemical theory and computation 2018, 14, 5645–5652.
  137. [141]Wang, Z.; Li, S.-S.; Wang, L.-W. Efficient real-time time-dependent density functional theory method and its application to a collision of an ion with a 2D material. Physical review letters 2015, 114, 063004.
  138. [142]Wigner, E. On the quantum correction for thermodynamic equilibrium. Physical review 1932, 40, 749.
  139. [143]Hillery, M.; O’Connell, R. F.; Scully, M. O.; Wigner, E. P. Distribution functions in physics: Fundamentals. Physics reports 1984, 106, 121–167.
  140. [144]Zobač, V.; Kuisma, M.; Larsen, A. H.; Rossi, T.; Susi, T. Ehrenfest dynamics with localized atomic-orbital basis sets within the projector augmented-wave method. The Journal of Chemical Physics 2025, 162.
  141. [145]Saunders, V. R.; Hillier, I. H. A “Level-Shifting” method for converging closed shell Hartree–Fock wave functions. International Journal of Quantum Chemistry 1973, 7, 699–705.
  142. [146]Schmerwitz, Y. L.; Levi, G.; Jónsson, H. Calculations of excited electronic states by converging on saddle points using generalized mode following. Journal of Chemical Theory and Computation 2023, 19, 3634–3651.
  143. [147]Adamo, C.; Barone, V. Toward reliable density functional methods without adjustable parameters: The PBE0 model. The Journal of chemical physics 1999, 110, 6158–6170.
  144. [148]Efron, B. Bootstrap methods: another look at the jackknife. The Annals of Statistics 1979, 7, 1–26.
  145. [149]Handt, J.; Kunert, T.; Schmidt, R. Fragmentation and cis–trans isomerization of diimide in fs laser-pulses. Chemical physics letters 2006, 428, 220–226.
  146. [150]Tuckerman, M.; Berne, B. J.; Martyna, G. J. Reversible multiple time scale molecular dynamics. The Journal of chemical physics 1992, 97, 1990–2001.
  147. [151]Isborn, C. M.; Li, X.; Tully, J. C. Time-dependent density functional theory Ehrenfest dynamics: Collisions between atomic oxygen and graphite clusters. The Journal of chemical physics 2007, 126, 134307.
  148. [152]Schleife, A.; Draeger, E. W.; Kanai, Y.; Correa, A. A. Plane-wave pseudopotential implementation of explicit integrators for time-dependent Kohn–Sham equations in large-scale simulations. The Journal of Chemical Physics 2012, 137, 22A546.
  149. [153]Lian, C.; Guan, M.; Hu, S.; Zhang, J.; Meng, S. Photoexcitation in Solids: First-Principles Quantum Simulations by Real-Time TDDFT. Advanced Theory and Simulations 2018, 1, 1800055.
  150. [154]Shepard, C.; Zhou, R.; Bost, J.; Carney, T. E.; Yao, Y.; Kanai, Y. Efficient exact exchange using Wannier functions and other related developments in planewave-pseudopotential implementation of RT-TDDFT. The Journal of Chemical Physics 2024, 161, 024111.
  151. [155]Jakowski, J.; Lu, W.; Briggs, E.; Lingerfelt, D.; Sumpter, B. G.; Ganesh, P.; Bernholc, J. Simulation of 24,000 Electron Dynamics: Real-Time Time-Dependent Density Functional Theory (TDDFT) with the Real-Space Multigrids (RMG). Journal of Chemical Theory and Computation 2025, 21, 1322–1339.

Paper details

vlnc:W_cm2wq9qb
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 6

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents