An Excitation Strategy for the Initial Condition Generation for Surface Hopping Trajectories Using Electron-Only Dynamics Including Explicit Laser Pulses Authors: Lorenz Grünewald, Laurens van Dam, Sebastian Mai Published: 2025 ABSTRACT Nonadiabatic dynamics simulations, e.g., via trajectory surface hopping, are nowadays used regularly to describe various photoinduced phenomena in molecules. For a number of reasons, in the setup of such simulations, the actual photoexcitation process is often described by rather crude and approximate excitation schemes (e.g., vertical excitation by an implicit delta pulse), and more attention is directed to the ensuing dynamics simulations. However, several studies have implied the importance of properly considering the spatial, temporal, and spectral details of the exciting laser pulse. Here, we suggest the “electron-only explicit” (“EOE”) excitation scheme for setting up trajectory surface hopping initial conditions based on an explicit laser pulse at little computational cost. The scheme is based on solving the time-dependent electronic Schrödinger equation including the explicit influence of a laser pulse within the frozen-nuclei approximation. The obtained time-dependent, coherently excited, electronic populations are then used to stochastically select the initial electronic states in a postprocessing step. Here, the electronic populations are renormalized such that one excites a reasonable fraction of the initial condition even when operating well within the weak-field regime. The new scheme is made freely available as part of the SHARC 4 dynamics package. We illustrate and validate the new excitation scheme by means of several simulations of sodium iodide and 6-cyanobenzquinuclidine excited with laser pulses of different energy and pulse duration, comparing to quantum dynamics results. Introduction The development of ultrafast laser spectroscopy has enabled trackingand potentially steeringmolecular dynamics on (sub)-femtosecond time scales. − Still, the interpretation of the molecular dynamics during and after photoexcitation is highly nontrivial and challenging due to the intrinsic interaction of all nuclear and electronic degrees of freedom. Nonadiabatic molecular dynamics (NAMD) simulations that specifically incorporate the photoexcitation process and the subsequent processes can significantly contribute in understanding ultrafast processes. One prominent example for such an NAMD method is the trajectory surface hopping (TSH) method. − Within this mixed quantum-classical method, a swarm of trajectories is propagated, in which nuclei follow classical trajectories driven by the gradient of a single adiabatic state that can stochastically switch (hop) based on the quantum mechanical evolution of the electronic wave function. To explicitly account for the photoexcitation process in TSH, in principle, one could start the trajectories in the ground state and incorporate the interaction between the molecule and a laser field in the dynamics − (we will refer to this as the “full explicit” excitation scheme below). However, this is rarely done because a considerable part of the computational effort would be wasted on propagating trajectories in the ground state that never get excited by realistically weak laser pulses. − Enhancing the number of excited trajectories and thus the dynamics efficiency by artificially increasing the laser intensity is reported to lead to spurious behavior because of strong-field and interference effects. , Moreover, especially for longer laser pulses, interference effects between the ground-state and excited wave functions break the underlying assumptions of TSH, leading to inaccurate results and making the “full explicit” scheme undesirable. , Instead of explicitly treating the light-matter interaction in the TSH simulations, one can approximate the photoexcitation process in a cheaper way by starting the trajectories in suitably selected initial excited states. ,− In most TSH simulations, this is done by invoking a sudden vertical excitation, − which we refer to as “vertical” excitation scheme in the following. It treats photoexcitation as an instantaneous event and thereby neglects the true spectral and temporal structure of the exciting laser pulse. Although such “vertical” excitation schemes often consider an energy window around the central pulse frequency to restrict the set of potential initial states, there is no standard criterion for the window extent and the allowed excitation region isto some extentarbitrarily chosen, e.g., via the window bandwidth. , While such a relatively crude treatment of photoexcitation may suffice in certain contexts and may describe photochemical processes reasonably welle.g., if the focus of the simulation is on photodynamics that happens in a statistical manner hundreds of femtoseconds after an excitation pulseit potentially fails for scenarios where the dynamics on the time scale of the laser pulse is relevant. Recently, Janoš et al. highlighted the importance of carefully selecting initial conditions for trajectory-based NAMD simulations and emphasized how the commonly used “vertical” excitation scheme can produce misleading results that do not take into account the laser properties. Hence, they suggested the promoted density approach (called the “PDA” scheme below) to incorporate the temporal and spectral profile of a laser pulse in the selection of the initially excited states. The “PDA” is theoretically well founded and computationally inexpensive, and can describe a wide variety of photoexcitation scenarios, including excitation by chirped laser pulses. However, there are some aspects that this method does not fully describe. For example, the available implementation does not currently consider the (static or time-dependent) polarization of the laser, although it could in principle be included in the algorithm. Moreover, the “PDA” implementation implicitly assumes the electric dipole approximation, which truncates the multipole expansion of the light-matter interaction after the first term, and a plane-wave description of light. Including higher-order terms can be of interest when considering short wavelengths, structured light, tightly focused laser beams, or other nonplane-wave scenarios. − Lastly, the “PDA” does not consider the simultaneous/competing electronic dynamics that might occur during the pulse duration due to other types of couplings besides light-matter interactions, for example, strong spin–orbit couplings that lead to few-fs intersystem crossing. However, the mentioned physical phenomenapolarization, structured light, and intersystem crossingcan be critical for the description of the photoinduced dynamics, e.g., in transition metal or lanthanide complexes. , In this work, we propose a complementary excitation scheme, the “electron-only explicit” (“EOE”) scheme, for preparing initial conditions for TSH simulations. The design goals of this scheme are (i) to include the effect of the exciting laser at the level of the initial conditions, (ii) to require only one single-point calculation per initial condition to have similar computational cost as other excitation schemes, (iii) to include laser fields with arbitrary polarization and potentially beyond the dipole approximation, and (iv) to consider that other couplings besides light–matter interactions could be active during the excitation. Our method is based on two steps, where first the photoexcitation is simulated explicitly by propagating the electronic wave function in the presence of a laser field for each initial condition, assuming frozen nuclei. In a second step, based on the evolution of the electronic wave function coefficients, the initial electronic states and starting times for each initial condition are selected stochastically, where the coefficients are renormalized to increase the number of accepted initial conditions. This renormalization solves the problem that otherwise weak laser fields would yield only very few excited initial conditions. Subsequently, the selected initial conditions are propagated with regular TSH, just like after using other excitation schemes. Our method is implemented in the SHARC 4 package and offers high computational efficiency due to the use of frozen nuclear coordinates during the laser interaction simulations. At the same time, it naturally incorporates laser beam polarization and is compatible with focused beams and nonplane-wave propagation regimes, − e.g., in scenarios where a nearly isolated magnetic field occurs. − While we discuss the “EOE” scheme in the context of TSH, we note that our proposed photoexcitation scheme can also be coupled to the coherent switching with decay of mixing (CSDM) approach, combining advantages of the Ehrenfest method and TSH. We demonstrate the validity and applicability of our method by analyzing the photodynamics of sodium iodide (NaI) and 6-cyanobenzquinuclidine (CBQ) excited by a variety of Gaussian laser pulses. Theory In this section, we start by briefly recapitulating the general framework of TSH and provide relevant equations. We then proceed to summarize the currently used photoexcitation schemes in TSH to provide context for the proposed “EOE” excitation scheme. Finally, we derive the key equations and parameters for the new excitation scheme. Surface Hopping The TSH approach denotes an NAMD method that treats the intertwined electronic and nuclear dynamics within a mixed quantum-classical picture. Due to its conceptual simplicity, efficiency, and accuracy, TSH is nowadays among the most widely used flavors to perform NAMD. The method considers the nuclear motion to follow classical Newtonian mechanics, while the electronic dynamics is described by a time-dependent electronic Schrödinger equation. The involved time-dependent electronic wave function is constructed as a linear combination of the eigenfunctions of the electronic Hamiltonian for the current nuclei configuration R⃗ |Ψ[R⃗(t),t]⟩=∑αcα(t)|ϕα[R⃗(t)]⟩, 1 in which α runs over all basis states in the set of included states, c α(t) are time-dependent complex coefficients, and |ϕα[R⃗(t)]⟩ are the electronic basis states. At this point, one is free to choose a basis for the wave function, e.g., a (quasi-) diabatic basis that is constant at all positions or the adiabatic basis consisting of the eigenfunctions of the electronic Hamiltonian Ĥ. We will comment below on the implications of this freedom of choosing the basis representation for the new excitation scheme. The wave packet expansion can be inserted into the time-dependent Schrödinger equation to obtain an expression for the electronic equation of motion 2 ∂cβ(t)∂t=−∑α[iHβα+Kβα]cα(t), where β is one electronic state, Hβα=⟨ϕβ|Ĥ|ϕα⟩ is a matrix element of the electronic Hamiltonian operator over two basis states, and Kβα=⟨ϕβ|∂∂t|ϕα⟩ is an element of the time-derivative coupling matrix. , Here, different choices for Ĥ are possible; the simplest includes only the electron kinetic energy and the intermolecular Coulomb interactions (the molecular Coulomb Hamiltonian, MCH). However, one can also include, e.g., spin–orbit couplings or light-matter interactions (which would make Ĥ time-dependent). The actual TSH simulation is based on an iterative algorithm, involving several steps. First, the nuclear positions R⃗ are updated based on the electronic forces. These forces are computed as gradients of the current “active state” β, which is one of the electronic basis states. The active state is determined stochastically from the evolution of the electronic coefficients c α of all states α. A change of active state is referred to as a “surface hop”. At the new nuclear positions, an electronic structure calculation is performed, the obtained gradients are used to update the velocities v⃗, and the energies and couplings are used to propagate the electronic coefficients using eq . Based on the fewest-switches surface hopping concept, several prescriptions to compute the hopping probabilities have been proposed. ,,, In this work, we adopt the global flux surface hopping (GFSH) scheme, which requires only the knowledge of the coefficients of the previous and current time steps, but not of any other intermediate. The GFSH hopping probability P β→α is computed as 3 Pβ→α(t+Δt)=(1−|cβ(t+Δt)|2|cβ(t)|2)×|cα(t+Δt)|2−|cα(t)|2∑γmax[0,|cγ(t+Δt)|2−|cγ(t)|2], where the sum in the denominator effectively goes only over all states γ whose population increases from t to t + Δt. The active state-active state probability P β→β as well as all negative probabilities P β→α < 0 are set to zero. In eq the first factor on the right-hand side is the overall probability to leave β according to the fewest-switches criterion, and the second factor splits this probability over the possible new states α. In each time step, after the hopping probabilities were computed from the coefficients, the new active state is stochastically selected by drawing a random number r(t) between 0 and 1. , A hop to a prospective new active state α is performed if 4 ∑γ=1α−1Pβ→γ(t)0,β>0 are anticipated to play only a negligible role in most cases, because the induced excited-state absorption is often nonresonant and scales nonlinearly with intensity, thus becoming negligible with the weak pulses applied in the “EOE” scheme. Thus, excited-state–excited-state transition dipole moments and the corresponding higher-lying electronic states can be neglected unless anticipated to play a role. Having the freedom to neglect the excited-state–excited state transition dipole moments is advantageous if the electronic structure method cannot provide these quantities (e.g., from some TD-DFT codes). Initial State Selection Process The electron-only dynamics simulations are carried out with the goal of obtaining the time evolution of the electronic populations |c α(t)|2 for each initial condition k. The core step in our new excitation scheme is obtaining the initial electronic state for k and the corresponding excitation time t′ for the subsequent nonadiabatic trajectories. This process involves two smaller steps, which are the renormalization of the population and the application of the surface hopping algorithm to the populations (as discussed below). This is the first step in the excitation scheme in which the choice of electronic representation matters. In SHARC, one commonly has two possible choices here. The first choice is the representation of the electronic states as obtained from the electronic structure, i.e., which are spin-free and not field-dressed (this basis is typically called the MCH representation within SHARC). The second choice is the representation of the eigenstates of the full electronic Hamiltonian (called the diagonal representation), which are spin-mixed if spin–orbit couplings are included and field-dressed if light–matter interactions are present. Here, it is important to realize that the subsequent TSH simulations will be performed without including the laser field. Hence, the diagonal representation is not identical in the electron-only simulations and the TSH simulations, such that the selected initial state would be inconsistent. Therefore, the initial state selection described in the following is carried out in the MCH representation (see also ref for further discussions). We note that the choice to select the initial state in the MCH representation is the standard approach in SHARC, even if the actual TSH simulations are carried out in the diagonal representation. Here, during the first time step of the TSH simulation, the active diagonal state is chosen as the one that most strongly overlaps with the selected MCH state. The initial coefficients of the selected MCH state are set to 1 and then transformed into the diagonal representation. − In this way it can be assured that at t = 0 the initial electronic wave function has a well-defined spin, which is important for simulating intersystem crossing. To find the largest acceptable renormalization constant for the electronic populations, we analyze the likelihood of exciting from the state β (the state in which the electron-only dynamics is initialized) to all other states (which are the potential initial states for the subsequent TSH simulation). The probability to leave β at time step t is given by 8 Pβ→any(t)=max(0,1−|cβ(t+Δt)|2|cβ(t)|2). From this, one can find the total probability to leave state β in any time step of the simulation as the complementary probability of remaining in β for all steps 9 Ptot,k=1−∏t(1−Pβ→any,k(t))=1−∏tmin(1,|cβ,k(t+Δt)|2|cβ,k(t)|2). We note that this P tot,k is equal to the sum of the population of the excited states after the pulse ended, but assuming that no population was transferred back to β (this is the effect of the min function). In order to have the maximum acceptance ratio in the excitation sampling, this probability should be as high as possible. However, there will be large differences in P tot,k across the different initial conditions kfor some k, the excitation energies/transition dipole moments will be in resonance/parallel with the laser field, favoring population transfer to the excited states. Meanwhile, for other k nearly no population transfer will take place. Thus, the optimal acceptance ratio will be reached if we take the global maximum of P tot,k over all k Pmax=maxkPtot,k. 10 This value quantifies the probability to get excited to any excited state for the initial condition k where this is most likely. This value P max is the sought-after renormalization constant. In the next step, we apply the renormalization constant to the time-dependent excited-state populations for all initial conditions k and all excited states α |c̃k,α(t)|2=|ck,α(t)|2Pmax,∀α≠β. 11 and rescale the population of β accordingly |c̃k,β(t)|2=1−∑α≠β|c̃k,α(t)|2=|ck,β(t)|2−(1−Pmax)Pmax. 12 Since P max will typically be a positive value close to zero, this renormalization significantly increases the population transfer to the excited states. We note that it implies for the differences between two time steps |c̃k,α(t+Δt)|2−|c̃k,α(t)|2=1Pmax(|ck,α(t+Δt)|2−|ck,α(t)|2) 13 for all α, even including α = β. Thus, we enhance population transfer between the states without otherwise changing the electronic dynamics. Alternatively, the renormalization procedure can be thought of as neglecting the nonexcited initial state population and normalizing the remainder for producing the hopping probabilities. In the next steps, we compute surface hopping probabilities from the renormalized populations |c̃k,α(t)|2 in order to stochastically find the initial excited state for the subsequent TSH simulations. Here, we use the GFSH formalism because it depends only on the populations but not on any other quantity that might not be available during postprocessing. We note that GFSH is independent of the phase of the coefficients, so we can apply the renormalization directly to the populations. Plugging the renormalized populations from eqs and into the GFSH probabilities in eq , we obtain 14 P̃β→α(t)=(1−|cβ(t+Δt)|2−(1−Pmax)|cβ(t)|2−(1−Pmax))×|cα(t+Δt)|2−|cα(t)|2∑γmax[0,|cγ(t+Δt)|2−|cγ(t)|2] for hopping probabilities from the initial state β. Note that only the prefactor in parentheses is affected by the renormalization, whereas the second factor stays the same as in eq . For hopping probabilities from an excited state, P̃α→β or P̃α→γ , the renormalization preserves the unmodified eq , because the 1Pmax factors in eqs and cancel out in both factors of eq . It can be shown by induction that 1 – P max ≤ |c β(t)|2 for all time steps and for all trajectories (see Section S2 in the Supporting Information). Consequently, the numerator and denominator of the first factor in eq cannot become negative, such that the renormalization as described does not affect the sign of the hopping probabilities. If the hopping probabilities from eq at a time step t are positive, they will also be positive when evaluated with eq . Furthermore, evaluating P̃β→α(t)Pβ→α(t) from eqs and leads to a factor of |cβ(t)|2|cβ(t)|2−(1−Pmax) , which for |cβ(t)|2≈1 (as is generally obtained with weak lasers) simplifies to 1Pmax . This is the factor by which the renormalization enhances the large majority of hopping probabilities out of the initial (lower) state β of all initial conditions. The hopping probabilities are enhanced by a larger factor if |cβ(t)|2 approaches 1 – P max. This makes sense if one considers that |cβ(t)|2→1−Pmax implies |c̃β(t)|2→0 , and that means that the hopping probabilities need to approach 100% to ensure the highest possible acceptance ratio of initial conditions. As a side remark, we note that the presented approach to derive enhanced hopping probabilities from the renormalized populations is preferable to a simple uniform rescaling of the original hopping probabilities. This is because a rescaling of the per-time-step hopping probabilities with a constant factor does not produce a uniform rescaling of the total excitation probabilities of each initial condition, as shown in Section S3. Once the hopping probabilities are computed via eq , we can obtain the initial electronic state α k (0) and starting time t k ′ for each initial condition k. Here, for each k, we simply apply the stochastic sampling from eq , starting at t = 0 and running over all time steps. If a hop is performed to a state α, the hopping probabilities for the remaining steps are recomputed for leaving α, so that hops back to the initial state and hops to other coupled states are possible. The state to which the last hop is performed is taken as α k (0) and the corresponding time as t k ′. If the last hop is to the initial state β, the initial condition is not selected, and no trajectory will be started from it. Similarly, initial conditions for which no hops are performed at all are also not considered further. We note that the inclusion of more than one hop per initial condition in the algorithm is based on practical considerations. Preliminary tests that were restricted to a maximum of one hop per initial condition showed that there is a significant number of off-resonance states that get selected unexpectedly. These states get temporarily populated by the interaction with the laser, but due to being off resonance, they are depopulated in the second half of the pulse, so that no net population remains. Admitting a second hop in the EOE selection algorithm ensures that these off-resonance states are removed from the set of selected initial states. Third hops as well as hops from the excited state to other excited states were exceedingly rare in the examples shown below. Such hops would otherwise introduce some inaccuracies in the description of the excitation, because the nuclei actually move between the different hops, which is not described by the frozen-nuclei EOE simulations. Once each initial condition k has been processed and the initial electronic state α k (0) and starting time t k ′ are obtained for each (or determined that k is not selected at all), this information is written to an output file, using the standard file format for SHARC initial conditions. As part of this work, the file format was slightly extended to allow storing the starting time t k ′. Here we note that we consider each initial condition k only once, ensures that each excited initial condition is fully statistically independent. Full Dynamics Simulation After the initial state selection process, the excitation-scheme-specific steps are finished. Subsequently, the actual TSH trajectories (with moving nuclei) are set up normally, using α k (0) as the initial electronic state. The initial electronic coefficients are set as cα,k(0)=δααk(0) in the MCH representation, as mentioned above. As the effect of the excitation laser has already been incorporated by means of the initial state, the TSH trajectories do not need to include the laser explicitly in the Hamiltonian. During the setup, the time of excitation t k ′ is written into a small file in each trajectory folder. Once all TSH trajectories are finished, the trajectories can be shifted in time according to t k ′ during postprocessing. This allows to recover the true temporal distribution of the excitation process for finite laser pulses. This time shifting aspect of our proposed excitation scheme is essentially the same as in ref . Computational Details To showcase the proposed “EOE” excitation scheme, we carried out several sets of simulations on the sodium iodide (NaI) system using the SHARC package , including the newly implemented “EOE” scheme. The nonadiabatic dynamics of NaI is well-known historically for its role in the development of ultrafast pump–probe spectroscopy and quantum dynamics simulations. − Moreover, it has been used in the previous decades as an important test system for various nonadiabatic molecular dynamics methods. ,− Here, we employ it due to its useful combination of nontrivial excited-state dynamics, the availability of analytical models for the potential energy surfaces, low dimensionality that makes exact quantum dynamics calculations feasible, and its use in previous studies investigating excitation schemes. , In brief (details below), based on a single set of initial geometries/velocities, we generate several different sets of initial conditions with differently obtained initial states based on different laser pulses. We use three excitation energies (low: 3.68 eV, mid: 3.89 eV, high: 4.15 eV) and three pulse durations (intensity FWHM’s of 20 fs, 100 fs, and 500 fs), equivalent to the test simulations for the “PDA” scheme in ref . For comparison purposes, we generate three additional sets of initial conditions with the “vertical” excitation scheme (low/mid/high excitation energy ±0.03 eV). From these 12 sets of initial conditions, we launched TSH simulations to obtain the time-dependent mean ⟨R⟩S1 and standard deviation ⟨ΔR⟩S1 of the Na–I distances. Reference results were obtained with nine grid-based quantum dynamics simulations (low/mid/high excitation energy, each with 20 fs/10 fs/500 fs pulses). Potential Energy Surfaces and Quantum Dynamics Simulations We used a diabatic two-state Hamiltonian reported in the 1970s and 1980s for NaI. , The corresponding equations are given in Section S4. The quantum dynamics simulations used a spatial grid of 1024 points between 3 and 40 Bohr. The ground state vibrational wave function was obtained via imaginary time propagation and served as initial wave function for the QD simulations. The dynamics simulations were carried out with the split-operator propagator, using a 0.1 fs time step. Convergence was checked against a calculation with 4096 grid points and 0.01 fs time steps; deviations in ⟨R⟩ and ⟨ΔR⟩ were <0.01 Å. Overall, we performed nine quantum dynamics simulations to serve as reference for the TSH simulations. Laser Pulses The nine laser pulses were generated based on this analytical functional form (see Section S5 in the Supporting Information ,, )­ E→(t)=E0·e−4ln⁡2FWHM2(t−tc)2·cos(ω0(t−tc)+θ)e⃗. 15 Here, E0 is the maximum electric field amplitude, FWHM is the full-width-at-half-maximum of the electric field amplitude/envelope, t c is the temporal center of the envelope, θ is the phase at t = t c, and e⃗ is the polarization unit vector. The pulses had central energies of 3.68 eV (called “low”), 3.89 eV (“mid”), and 4.15 eV (“high”). The intensity FWHM’s were 20 fs, 100 fs, and 500 fs. The maximum electric field strength was set to 0.001 au and the phase was θ = 0. The center of the pulses was shifted to t = 0 in all figures below. The polarization vector was chosen parallel to the NaI molecular axis. Phase Space Sampling and Spectrum We sampled 10,000 initial conditions from the numerically computed Wigner distribution , of the ground state vibrational wave function (i.e., at 0 K) obtained from imaginary time propagation as mentioned above. All subsequent calculations were based on this one set. For each geometry, a single-point calculation was carried out using the analytical model. From this data, we simulated the absorption spectrum as a Gaussian convolution with full width at half-maximum (FWHM) of 0.1 eV. The spectrum is shown in Figure . Simulated absorption spectrum of NaI computed from 10,000 Wigner samples using the analytical model. The dashed lines indicate the central energies of the low-, mid-, and high-energy laser pulses. Initial State Selection Initial conditions with the “vertical excitation” scheme were selected from three energy windows, 3.65–3.71 eV, 3.86–3.92 eV, and 4.12–4.18 eV, using selection probabilities for each electronic state proportional to foscΔE2 . For the initial conditions prepared with the “EOE” scheme, each of the nine laser pulses was used. The electron-only simulations were carried out with a 0.0625 fs time step and propagated until the electric field was effectively zero. The laser polarization was not randomized. As described above, in these simulations, nuclear motion was ignored, the same electronic energies and dipole moments reused in all time steps, and hopping and decoherence corrections were turned off. From the results of these trajectories, the initial electronic state and starting time were obtained stochastically (based on eq ) for all 10,000 initial conditions. We note that only every eighth step of the electron-only simulations was written out, making the selected start times multiples of 0.5 fs, i.e., multiples of the time step of the subsequent TSH simulations. The results of the initial state selection was analyzed by means of scatter plots of the excitation energy and excitation time of the accepted initial states, and a comparison of the scatter plots with the pulses’ Wigner distribution. Nonadiabatic Dynamics Simulations Based on the 12 sets of initial conditions (three from “vertical” excitation, nine from “EOE”), we launched TSH simulations, including two states. The nuclear time step was 0.5 fs, and the electronic wave function was propagated with 0.02 fs steps using local diabatization, using wave function overlaps computed from the adiabatic-diabatic transformation matrices at subsequent time steps. The TSH trajectories were propagated for between 1600 and 2200 fs, depending on the length of the laser pulses. No explicit laser was included here. The simulations used SHARC surface hopping probabilities (rather than GFSH probabilities that are used within the “EOE” scheme), an energy-based decoherence correction with the recommended decoherence parameter value of 0.1 a.u., and rescaling of the full velocity vectors along their direction. ,,, Trajectories were analyzed by computing the time-dependent average and standard deviation of the Na–I distance for each swarm. Results and Discussion In the following, we discuss the selection of the initial electronic states in the employed excitation schemes, the obtained nonadiabatic dynamics of NaI, as well as the agreement with the quantum dynamics reference calculations and the “PDA” excitation scheme. Excited-State Selection We first discuss the energetic and temporal distribution of the selected initial conditions from the “EOE” scheme for the low-energy, 20 fs pulse. Here, we compare our distribution to one obtained with the “PDA” scheme, because the “vertical” scheme cannot be used to compare starting times. The comparison between “EOE” and “PDA” is given in Figure . In panel a, we show the simulated absorption spectrum of NaI, the intensity spectrum of the employed pulse, and histograms of the distribution of excitation energies of selected initial states. We can see that “EOE” and “PDA” give consistent results. Both select initial states only from within the laser’s energetic spectrum, as expected. The most likely excitation energy agrees approximately with the central energy of the laser. Furthermore, the histograms show a stronger tail at high energies, consistent with the absorption spectrum of NaI that shows higher intensity at energies above the laser central energy. Given that we used a constant diabatic transition dipole moment, this higher spectral intensity arises from a higher occurrence of large excitation energies in the Wigner distribution of the ground state. Comparison of the distribution of excitation energies and excitation times for the low-energy, 20 fs laser pulse between the “EOE” and “PDA” schemes. Panel (a) shows the absorption spectrum of NaI (blue, from Figure ), the pulse spectra intensity (gray), and histograms of the excitation energies of the selected initial states. Panel (b) shows the temporal intensity envelope of the pulse and histograms of the excitation times of the selected initial states. Panel (c) shows scatter plots of the excitation energies and times for the two methods, together with the Wigner distribution of the laser pulse (note the logarithmic color scale). Panel (d) shows the pulse’s electric field over time for reference. In panel b, we show the temporal envelope of the pulse together with histograms of the excitation time. Both excitation schemes produce similar results, although the excitation probability before the pulse center is somewhat reduced for the “EOE” scheme. The scatter plots in panel c likewise indicate a similar performance of the “EOE” and “PDA” schemes; however, the former scheme produces a triangle-shaped tail for on-resonance transitions. We discuss these observed differences in more detail below. We note that in the “EOE” scheme, we only pick excitation times that are a multiple of the time step of the subsequent TSH simulations, to aid in the statistical analysis of the time-shifted trajectories. This is the reason that the “EOE” results in panel c exhibit a discrete distribution of starting times. Overall, both excitation schemes produce consistent excitation distributions, indicating that the “EOE” scheme correctly describes the excitation process and the nuclear density of the excited/promoted wave function. In Figure , we present additionally all excitation energy–excitation time scatter plots for all nine laser pulses, each computed with the “EOE” scheme. The labels in the corner of each panel indicate the number of initial conditions that were selected for each laser pulse. For reference, using the “vertical” excitation scheme, 848, 1356, and 474 initial conditions were selected for the low-, mid-, and high-energy selection windows (with windows of ±0.03 eV width, i.e., similar to the 20 fs pulse). It can be noted that the EOE scheme with the 20 fs pulse produces the largest number of selected initial states, compared to the 100 and 500 fs pulses. This can be explained by the spectral bandwidth of the different pulses. The number of selected initial states is reduced by approximately a factor of 5 between the 20 and 100 fs pulses (and the 100 and 500 fs pulses), in accordance with the inverse proportionality of spectral bandwidth and temporal width. Comparison of the Wigner distributions of the nine used laser pulses with the energetic and temporal distribution of the selected initial conditions from the “EOE” selection scheme. Note that each column of panels uses the same time axis, while the energy axes are different for each panel. The labels give the number of selected initial states, based on a set of 10,000 initial geometries/velocities. Besides the number of selected initial conditions, Figure shows the distribution of excitation energy and starting time. It can be seen that, in each case, the excited initial states generally fall within the Wigner distribution of the pulse. However, especially for the panels with a large number of selected initial conditions (panels a,d), the scatter plot shows a teardrop-shaped distribution, with a reduced selection probability at early times and a triangular tail at late times. This triangular tail also indicates that, for off-resonance transitions, essentially no starting times near the end of the pulse were obtained. The teardrop-shaped distribution of selected initial conditions can be explained by an analytical model of the excitation process that takes place within each electron-only simulation. We assume coherent excitation of a static two-state system, using first-order perturbation theory, the rotating wave approximation, a pulse of the form E(t)=E0e−t2/2τ2cos(ω0t) , and a unit coefficient in the ground state before the pulse. It can be shown that the coefficient of the upper state is proportional to the integral over the pulse envelope, which for a Gaussian pulse relates to the (complex) error function. , The electronic coefficient of the upper state is cα(t)=−2πiμβαE0τ4ℏe−τ2Ω2/2(1+erf(t2τ−iτΩ2)) , where Ω = ω0 – ωβα is the detuning of the laser from the transition β → α. Computing the time derivative of the population |c α|2 gives ∂∂t|cα|2(t;Ω)=2πμβα2E02τ4ℏ2e−t22τ2−τ2Ω22×R[eitΩ(1+erf(t2τ−iτΩ2))]. 16 For the two-state case, |c β(t)|2 = 1 – |c α(t)|2, hence the derivative in eq is formally proportional to the hopping probabilities in eq (with negative probabilities set to zero). Figure plots eq together with the marginal probabilities in time and frequency. It can be observed that the shape of the distribution only depends on τ, which simply stretches/compresses it; hence, we plot eq in reduced coordinates tτ and τ·Ω. The figure reveals that the analytical model in eq produces a teardrop-shaped distribution of the excitation probability that is fully consistent with the scatter plots in Figures and . Minor differences between the model in eq and actual distributions from the “EOE” scheme can possibly arise from the accounting for additional electronic states, second-order coupling between excited states, and geometry-dependent transition moments that are not considered by eq . Figure raises several implications for the excitation probability distributions. The most probable excitation time (black dots) depends on the detuning of the respective transition from the central laser frequency: on-resonance transitions tend to get excited slightly after the pulse center. For this reason, the temporal marginal probability (panel a) is slightly shifted to later times compared to the pulse envelope. This is because for on-resonance transitions the population grows with the square of the accumulated pulse intensity over time, i.e., follow a |1 + erf­(t)|2 behavior. In contrast, the frequency marginal probability (panel b) exactly matches the pulse’s power spectrum. Plot of eq , the temporal derivative of the upper state population for the coherent excitation in a two-level system, with reduced time t/τ and reduced energy detuning τ·Ω. Panels (a,b) show the time and frequency marginal (i.e., integrated) probabilities, while panel c shows the full distribution. In panel (c) black dots mark the time of maximum excitation probability for each detuning. The analytical model illustrates that the observed distribution of initial conditions is a consequence of treating the excitation process fully coherently and with a constant transition energy for each initial condition in the “EOE” scheme. This coherent treatment explains the differences between the excitation distributions in Figure and the pulse Wigner distribution in Figure . Whether the assumption of a fully coherent excitations is justified depends on the molecular system under investigation. We expect thus that the “EOE” scheme works best for systems where excitation energies are only weakly geometry-dependent in the Franck–Condon region, like in f–f transitions in lanthanide complexes. For systems where the assumption of coherent excitation is not reasonable, we anticipate that adding a decoherence correction to the “EOE” scheme would produce more time-localized excitation rates. In such a scheme, the distribution of excitation time and energy would be more similar to the Wigner distribution of the pulse and, consequently, more comparable to the results of the PDA scheme. Nonadiabatic Dynamics The main results of the nonadiabatic dynamics simulations and quantum dynamics simulations are shown in Figure , which presents the expectation values (panels a–d) and standard deviations (panels e–h) of the swarm of trajectories (or wave packets). Plots of the time-dependent nuclear densities from the TSH and quantum dynamics simulations are given in Figure S1 in Section S6 of the Supporting Information. In Figure , panels a and e show the results for the TSH simulations launched from the “vertical” initial conditions in three different energy windows (the width of the window corresponds to the width of the 20 fs pulse). It can be seen that already this rather crude excitation scheme can correctly describe the evolution of the bond length expectation value. However, as shown in panel e, the swarms of trajectories are significantly too narrow, compared to the quantum dynamics reference (dashed lines). Comparison of the temporal evolution of the expectation values ⟨R⟩S1 (a–d) and standard deviation ⟨ΔR⟩S1=⟨R2⟩S1−⟨R⟩S12 (e–h) of the Na–I bond length. We show results for the “vertical” excitation scheme (a,e) and the “EOE” scheme with 20 fs (b,f), 100 fs (c,g), and 500 fs (d,h) pulses. Dashed lines show the reference quantum dynamics (QD) results, and full lines plus shaded areas represent the TSH results. The dashed lines in a,e were taken from the 20 fs pulse quantum dynamics simulation. Corresponding nuclear density plots are given in Figure S1. The results obtained with the “EOE” scheme provide generally a better agreement with the reference results. In panels b,f, the results for the 20 fs pulse are shown. While the “EOE” scheme improves the bond length expectation values only moderately over the “vertical” scheme, the improvement for the standard deviations is very notable. Thus, similarly accurate results as with the “PDA” scheme are obtained for short laser pulses. For the somewhat longer 100 fs pulse, the “EOE” scheme still produces reasonable results (again of similar accuracy as the PDA scheme), as evidenced in panels c and g. Finally, the results for the long 500 fs pulse in panels d and h clearly show the limitations of the excitation scheme, even though some qualitative agreement is still achieved for the high-energy pulse (see Figure S1 for why the agreement is only qualitative). However, we note that sodium iodide is a system with rather heavy atoms and rather flat potential energy surfaces, where the electronic energies are not expected to vary very quickly, so that longer pulses might lead to smaller deviations than in other systems. Further applications of the scheme to other molecular systems, e.g., some of the benchmark systems used in the nonadiabatic dynamics community, , will show under which conditions the “EOE” scheme provides the most robust results. Application to a Molecular System To show that the “EOE” scheme is also applicable to polyatomic systems, in Section S7 and Figure S2 we present simulations on the CBQ molecule, a rigid analogue of the “molecular Tully II model” 4-(N,N-dimethylamino)­benzonitrile. CBQ is described by a 5-state linear vibronic coupling model with full rotational and translational invariance at the TDA-ωB97X-D/def2-TZVP level of theory. 10,000 Wigner samples were generated and excited with either the “vertical excitation” scheme (5.95–6.05 eV window) or the “EOE” scheme with a laser pulse with 6.0 eV central energy and 18.2 fs temporal FWHM (corresponding to a 0.1 eV spectral FWHM). The selected initial conditions were propagated with SHARC for 120 fs. The results in Figure S2a–c show strong population oscillations (20 fs period) between the adiabatic S3 and S4, which, however, are mostly washed out when applying the time shifting described in Section . The peak population of the S3 is reduced from over 75% to about 50%, highlighting the importance of explicitly including the photoexcitation process when comparing to experiment. Additionally, we show in Figure S2d–f that the “EOE” scheme can be used to prepare initial conditions including a realistic distribution of partially aligned molecules with respect to a linearly polarized pump laser. Such initial conditions can be used to, e.g., simulate polarization-dependent scattering experiments. Conclusions We have presented a scheme to sample the initial electronic state for TSH simulations according to the interaction with an external laser pulse. The scheme is based on electron-only quantum dynamics simulations (i.e., for frozen nuclear coordinates) that explicitly include the relevant light–matter interaction Hamiltonian and can also include other contributions to the electronic Hamiltonian. Once the electron-only dynamics simulations are finished, the evolution of the electronic populations is renormalized (to increase the acceptance ratio) and used to derive surface hopping probabilities, which are used to stochastically identify the initial electronic state and start time for each geometry. The only required input are the sampled initial geometries, excitation energies and transition dipole moments at these geometries, and a laser pulse. This scheme is effectively an approximation to TSH simulations that explicitly include the laser pulse in the propagation, but the new scheme is much more economic and about as affordable as other commonly used excitation schemes. , The main advantages of the scheme are (i) that it can include various waveforms, (ii) that arbitrary interactions can be included in the electronic Hamiltonian (e.g., beyond-electric-dipole interactions relevant for nonplane-wave beams or high photon energies, spin–orbit couplings), and (iii) that initial conditions with polarization-induced anisotropy can be created. As it is based on the coherent propagation of the electronic coefficients, it is most appropriate for molecular systems with slow Franck–Condon motion after excitation, but should work generally well for sufficiently short pulses. We have demonstrated, based on TSH simulations of sodium iodide, that the “EOE” scheme can accurately describe the excitation and impart the effect of specific laser pulses on the nuclear dynamics, with good agreement with quantum dynamics simulations. In a second example on the CBQ molecule, we have shown that the “EOE” scheme offers a way to include temporal broadening in the results of TSH simulations, and that it can be used to prepare initial conditions that are partially aligned to a linearly polarized pump laser. The examples also showed that the proposed excitation scheme is computationally very cheap. The frozen-nuclei SHARC simulations take only a few seconds for each initial conditions, thanks to the efficient PySHARC implementation, , and the excited-state selection scheme likewise provides little overhead. The code used in this work is available under the GNU Public License as part of the SHARC 4.0.2 release, and it will be fully integrated and documented in the upcoming SHARC 4.1 package. Supplementary Material