Part I identified a precise restriction: when every nuclear configuration retains the primitive translations, different primitive crystal momenta remain separate. The real-space nonadiabatic molecular dynamics (NAMD) route relaxes that restriction by enlarging the simulation cell. The larger cell admits more spatial patterns of atomic motion and contains the corresponding folded electronic states. Wavefunction changes along this trajectory supply nonadiabatic coupling (NAC), which enters electronic propagation and a surface-hopping description.

The sequence matters. Folding tells us which wavelengths and electronic components fit inside the calculation. Molecular dynamics (MD) supplies their actual motion. Electronic-structure snapshots supply energies and wavefunctions. Their time dependence supplies NAC. Finally, fewest-switches surface hopping (FSSH) converts electronic population flow into stochastic changes of an active-state label. This article follows that chain, distinguishes a prescribed nuclear path from FSSH with nuclear feedback, and connects it to the authors' official MoS₂/WS₂ tutorial results.

1. Enlarge the cell to admit more wavelengths

Consider a one-dimensional primitive lattice with spacing \(a\) and reciprocal vector \(b=2\pi/a\). An \(M\)-cell supercell has length \(L=Ma\) and reciprocal spacing \(B_s=b/M\). A displacement pattern with primitive wavevector \(q\) is compatible with this supercell when

\[ u_{l+M,\kappa}=u_{l\kappa}, \qquad e^{iqMa}=1, \qquad q_r=\frac{r}{M}b,\quad r=0,\ldots,M-1. \]

All these \(q_r\) differ by supercell reciprocal vectors, so they fold to supercell phonon wavevector \(Q=0\), or supercell \(\Gamma\). Relative to the primitive cell, most of them have nonzero wavevector. The three-dimensional condition is the same: the phase must be one under every supercell translation. For a diagonal \(M_1\times M_2\times M_3\) expansion, the compatible primitive q mesh has those subdivisions along the three reciprocal directions.

The electronic folding follows the same geometry. At supercell electronic momentum \(K=0\), a perfectly repeated equilibrium crystal contains the primitive components

\[ k_r=\frac{r}{M}b,\qquad k_r\equiv0\pmod{B_s}. \]

Each primitive band supplies \(M\) folded supercell bands. Thus a larger cell can include several primitive electronic momenta in one supercell momentum sector. This is a reorganization of the same states at the undistorted geometry; simply copying a perfect primitive crystal does not create a new coupling. Adding more supercell electronic K points samples additional sectors, but a trajectory that preserves the supercell translations still cannot couple different supercell K sectors.

2. Choose a supercell for the electronic valleys

The authors' official tutorial makes this choice concrete for a MoS₂/WS₂ heterostructure. The band-edge valleys include primitive \(K\), so the supercell must fold that point into the sampled supercell sector. For a hexagonal expansion, the tutorial requires \(3n\times3n\); it also discusses a rectangular alternative. A larger atom count alone does not establish that the desired valley has been included.

The archived example's production-MD structure is a rectangular 108-atom cell containing 18 Mo, 18 W, and 72 S atoms. Its internal atoms can move differently during MD. Such motion breaks the primitive translations while preserving the supercell boundary condition, allowing the folded primitive components to mix. The official electronic results are shown in Section 10, after the steps needed to obtain them.

3. MD supplies the motion inside the supercell

Once the cell is enlarged, its primitive copies must be allowed to have different displacements. The atoms still interact; the change is that the internal copies are no longer constrained to move identically. If every copy is constrained to follow the same displacement, the primitive translations remain intact and Part I's restriction survives. Increasing the number of atoms while retaining that constraint gains no finite-q nuclear motion.

In a prescribed-path calculation, one can equilibrate the chosen supercell at the target temperature and then record a production trajectory \(\mathbf R(t_j)\). The NVT-to-NVE sequence in Part 0 is one practical route: NVT controls the temperature during equilibration, and NVE production evolves at fixed particle number, volume, and total energy. This preparation is distinct from the later electronic dynamics.

Forces evaluated along MD need not be harmonic. The trajectory can include anharmonic motion within the spatial degrees of freedom of the supercell. The cell still excludes incommensurate wavelengths and imposes a longest available wavelength. Finite size, trajectory duration, and thermal sampling therefore remain physical convergence questions. A particular phonon need not be assigned to every frame in order to compute the wavefunction changes.

4. Each configuration defines an instantaneous electronic basis

At each saved geometry, solve the electronic-structure problem and retain a consistent band window. In a Kohn-Sham (KS) description, write

\[ \hat H_{\rm KS}(\mathbf R(t_j))|\phi_\alpha(t_j)\rangle =\epsilon_\alpha(t_j)|\phi_\alpha(t_j)\rangle. \]

The label \(\alpha\) identifies an instantaneous supercell state. At the perfect reference structure its primitive momentum content can be identified by folding. At a thermally distorted configuration, primitive translation symmetry is generally broken, so a state can contain several primitive momentum components. It is generally incorrect to assign every distorted eigenstate one exact primitive k label.

The real-space route obtains the time dependence directly from these configurations. It does not require a first-order Taylor expansion of the electronic Hamiltonian about the equilibrium structure. Its approximations instead include the chosen electronic-structure model, cell size, nuclear dynamics, retained bands, and subsequent dynamical closure.

5. Consecutive wavefunctions supply NAC

Use the scalar time-derivative convention of Part I:

\[ d_{\alpha\beta}(t)= \langle\phi_\alpha(t)|\partial_t\phi_\beta(t)\rangle =\dot{\mathbf R}\cdot\mathbf D_{\alpha\beta}, \qquad \mathbf D_{\alpha\beta}= \langle\phi_\alpha|\nabla_{\mathbf R}\phi_\beta\rangle. \]

The scalar \(d\) has units of inverse time; \(\mathbf D\) is the derivative-coupling vector. Where analytic derivatives are unavailable, adjacent-frame wavefunctions provide the overlap matrix

\[ S_{\alpha\beta}(t,t+\Delta t) =\langle\phi_\alpha(t)|\phi_\beta(t+\Delta t)\rangle. \]

After consistent phase and state matching, an antisymmetrized finite-difference estimate is

\[ d_{\alpha\beta}\!\left(t+\frac{\Delta t}{2}\right) \approx \frac{S_{\alpha\beta}(t,t+\Delta t) -S_{\beta\alpha}^{*}(t,t+\Delta t)} {2\Delta t}. \]

The NAC/Hammes-Schiffer–Tully discussion derives the general overlap construction. In a supercell, both wavefunctions belong to the same supercell momentum sector, while their primitive momentum content may change. The vanishing cross-sector overlap in Part I no longer excludes coupling between those folded components.

Arbitrary orbital phases and swaps of nearly degenerate eigenvectors can produce misleading overlap changes. Phase matching, subspace tracking, a sufficient band window, and timestep convergence are essential. A band index alone is not a stable state identity across snapshots. At degeneracies, retain and track the coupled subspace rather than dividing by an energy gap. If tracking rotates a nondegenerate set of eigenstates by a general unitary matrix, transform the energy matrix consistently; it will no longer necessarily be diagonal.

6. Energies and NAC form the electronic propagation matrix

Expand the carrier wavefunction in the moving adiabatic KS basis:

\[ |\Psi(t)\rangle=\sum_\alpha c_\alpha(t)|\phi_\alpha(t)\rangle. \]

The basis itself changes with time. Inserting this expansion into the electronic Schrödinger equation gives

\[ i\hbar\dot c_\alpha =\epsilon_\alpha c_\alpha -i\hbar\sum_\beta d_{\alpha\beta}c_\beta, \qquad H^{\rm eff}_{\alpha\beta} =\epsilon_\alpha\delta_{\alpha\beta}-i\hbar d_{\alpha\beta}. \]

The diagonal energies accumulate phase; the moving-basis term couples amplitudes. Since \(d_{\alpha\beta}=-d_{\beta\alpha}^{*}\), the effective matrix is Hermitian when its inputs are consistent. The energy difference alone cannot transfer population. Even very close states remain uncoupled if the relevant NAC is zero.

For an overlap-based implementation, the MD snapshot interval, electronic propagation step, and interpolation of energies and NAC are separate numerical choices. Norm conservation and anti-Hermiticity checks detect numerical problems, but neither establishes convergence of a carrier relaxation time.

7. FSSH samples the electronic population flow

Electronic amplitudes and the active state have different roles. The amplitudes describe a coherent superposition. FSSH assigns each trajectory an active-state label and uses stochastic switches to follow population transfer. With \(P_a=|c_a|^2\), the preceding coefficient equation implies

\[ \dot P_a=-\sum_{b\ne a}J_{a\to b},\qquad J_{a\to b}=2\,{\rm Re}\!\left[c_a^*d_{ab}c_b\right], \qquad J_{b\to a}=-J_{a\to b}. \]

Positive \(J_{a\to b}\) is outward flow from state \(a\). For an active trajectory on \(a\) with \(P_a>0\), the elementary fewest-switches attempt probability over a sufficiently small electronic step \(\delta t\) is

\[ g_{a\to b}= \frac{\delta t}{P_a}\max(0,J_{a\to b}),\qquad b\ne a. \]

A uniform random number selects a destination from the cumulative probabilities or leaves the active state unchanged. The step must be small enough for a valid total probability; a total above one calls for a smaller step or appropriate substepping. The full FSSH derivation and Tully benchmark give the general background. Here its connection to the supercell calculation is explicit: the wavefunction overlaps determine \(d\), which determines amplitude flow, which supplies the attempted switches.

A nonzero NAC does not guarantee a hop at each instant: electronic coherence and the direction of population flow also enter. Conversely, FSSH cannot generate a physically justified transition from a missing symmetry-forbidden coupling. Folding and nuclear sampling must provide the relevant degrees of freedom before the hopping algorithm is applied.

8. Specify whether the nuclei respond to a hop

In full FSSH with nuclear feedback, the active many-electron adiabatic surface supplies the force. Writing its total potential energy as \(U_A(\mathbf R)\), one propagates nuclear momenta using \(-\nabla_{\mathbf R}U_A\). An accepted switch to \(B\) adjusts the momenta to satisfy

\[ \sum_I\frac{|\mathbf P_I'|^2}{2M_I}+U_B(\mathbf R) =\sum_I\frac{|\mathbf P_I|^2}{2M_I}+U_A(\mathbf R). \]

For real derivative couplings, the standard adjustment changes momenta along the derivative-coupling vector; the corresponding velocity change includes the inverse nuclear masses. A scalar time-derivative coupling alone does not determine this multidimensional direction. If insufficient kinetic energy prevents an upward switch, it is frustrated. The total surface energy \(U_A\) must not be identified automatically with a single KS orbital energy \(\epsilon_\alpha\). A feedback calculation needs a consistent excited-state energy and force model.

A common real-space HefeiNAMD strategy instead uses the classical-path approximation (CPA): generate a nuclear trajectory in advance and reuse it for the selected carrier's electronic dynamics. In this route, active-state switches do not recompute the saved nuclear path or apply the full feedback loop above. This is the prescribed-path workflow described in the real-space methodology of Zheng et al., Eqs. (1)–(4).

The population-flow formula explains the basic FSSH idea; a particular program can additionally use thermal acceptance factors or decoherence corrections. Those choices must be stated with the actual version and protocol. They are separate from supercell folding and the overlap-based NAC construction. An ensemble of active-state occupations is also a distinct estimator from the ensemble average of \(|c_\alpha|^2\).

9. The practical real-space calculation chain

  1. Select the supercell. Relate its compatible primitive q and folded k components to the transitions of interest. Distinguish cell-size convergence from electronic integration-mesh convergence.
  2. Sample nuclear motion. Relax and equilibrate the supercell, allow the internal cells to move independently, and save the production trajectory with a documented ensemble and temperature.
  3. Prepare electronic snapshots. Use consistent electronic settings and retain the states needed to cover the initial excitation and accessible final energies.
  4. Construct energies and NAC. Match states and phases between adjacent frames; check the snapshot interval and the retained subspace.
  5. Propagate amplitudes and switches. Specify the initial electronic state, propagation step, hopping rule, decoherence treatment, nuclear feedback or CPA, and statistical sampling.
  6. Define the observable. Report whether a population means coherent electronic weight or active-state occupation. Momentum-resolved results require projection or unfolding onto primitive momentum sectors; supercell band labels alone do not supply them.

The resulting picture connects the steps without attributing extra physics to folding: the larger cell admits the motion, the trajectory changes the instantaneous states, NAC measures their rotation, and FSSH supplies a stochastic active-state description. Finite size, classical nuclei, electronic-structure errors, state truncation, and the chosen feedback/decoherence approximation still bound the calculation.

10. The authors' MoS₂/WS₂ tutorial results

The official example studies an electron initially associated with the WS₂ conduction states and its relaxation in a MoS₂/WS₂ heterostructure. It follows the chain above: ground-state MD, electronic snapshots, energies and NAC, then prescribed-path surface hopping. The two figures below are the authors' original images, reproduced without modification from Qijing Zheng's Hefei-NAMD Training, Figures 8 and 11, under the page's CC BY 4.0 license. The linked repository copies are fixed at commit b7c8fc5.

The input packaged with the results uses 2,000 snapshots separated by 1 fs, TEMP=300 K, bands 325–335, 100 initial conditions, 1,000 fs of electronic dynamics, and 20,000 hopping trajectories per initial condition. Its INICON selects different trajectory starting points and initial bands 327–331. The online explanation instead displays an illustrative 5,000-trajectory input and repeated band 329. These are distinct source records; the archived controls should be consulted when working with the archived data. Here we present the published result, without claiming a new execution or independent reproduction.

Official MoS2/WS2 tutorial figure: time-dependent Kohn-Sham energies with red WS2 and blue MoS2 layer character.
Original KS-energy figure by Qijing Zheng and the tutorial contributors (source image). The horizontal axis is MD time in fs and the vertical axis is energy in eV with the author's original reference. Red identifies WS₂ character and blue MoS₂ character, as labelled in the color bar. The curves show fluctuating instantaneous electronic levels and their layer character. Their colors describe wavefunction projections, rather than the population of a propagated electron. Click the figure for its full resolution.

This first figure illustrates the snapshot stage: moving nuclei change both the eigenvalues and the character of the electronic states. Population transfer additionally requires the wavefunction changes between snapshots and the electronic dynamics. The next figure shows that later stage.

Official electron-relaxation result: population moves toward lower-energy states while the dashed blue mean electron energy decreases over 1000 fs.
Original carrier-dynamics result from the same tutorial (source image). Time is in fs and energy in eV. The white–yellow–red–black scale encodes band populations from 0 to 1; the dashed blue curve is the mean electron energy. Each figure retains the author's original energy scale; interpret energy changes within each panel. Population accumulates at lower energies as the mean energy falls. The color scale here represents population, whereas the preceding figure uses color for layer character.

The author's accompanying mean-energy data decreases from 0.9038 eV at the first printed time, 1 fs, to 0.6141 eV at 1,000 fs. These values quantify energy relaxation in this example. A quantitative interlayer transfer curve would additionally combine populations with the layer projections. A momentum-resolved transfer curve would require primitive-momentum projection or unfolding.

For the analysis code, see the author's KS-energy and layer-projection script and population and energy-analysis script. The latter reads SHPROP.* active-state populations. The tutorial separately identifies PSICT.* as coefficient output. Keeping these quantities distinct connects the displayed result to the propagation and switching steps discussed above.

Sources and further reading