Exact Center-of-Mass Dynamics for Neural Real-Time TDDFT

Neural networks can now replace real-time time-dependent density functional theory (TDDFT). That is the method chemists use to watch the electrons of a molecule respond, step by step, to a laser pulse. On small molecules, a learned propagator turns hours of CPU time into seconds on a GPU. These surrogates are trained to imitate trajectories, though, not to obey the laws those trajectories satisfy. One such law is exact and cheap to evaluate. The total electronic momentum changes only because of the external field and the attraction of the nuclei. The electrons' mutual repulsion and the exchange-correlation term (the quantum correction for electron-electron effects that DFT approximates) exert zero net force on the electron cloud. A surrogate that breaks this law accumulates errors in the induced dipole, which is the quantity experimentalists actually measure. The error grows worst at long times, exactly where the surrogate is supposed to save the most compute.

We propose to build this law into the network rather than hope it learns it. The idea comes from soliton theory. In the forced soliton equation, a localized wave is pushed by an external source. Its state is split, by a change of variables, into a center-of-mass position and momentum that obey Newton's second law exactly, plus internal "radiation" modes that carry everything else. We apply the same split to the Kohn-Sham electrons (the fictitious independent electrons whose density matches the real one). The collective electronic momentum is advanced by an exact, density-dependent force law, and only the internal deformation is left to the network. In two sentences: a learned TDDFT propagator should get the electron cloud's center-of-mass motion right by construction, because that motion follows a known force law. We take the change of variables that makes this split natural from the forced-soliton literature, and we test whether it fixes long-horizon dipole and spectrum errors that soft penalties and more data do not.

Core Claim

Forced Soliton Equation and Semiclassical Soliton Form Factors (also Accelerating solitons) treats a kink soliton in a two-dimensional field theory. An external source accelerates it, and the soliton is analyzed in phase space. The authors canonically transform to the soliton's position and momentum $P(t)$ plus field fluctuations, and find that the source simply enforces $\dot P = F(t)$. All the hard dynamics lives in the internal sector, and they admit that sector has no known explicit time-dependent solutions. The same paper therefore supplies both the method and its limitation: the collective sector is exactly solvable while the internal sector is not.

That is precisely the division of labor we want between physics and a neural network. Kohn-Sham electrons in a molecule are not a soliton, and we make no claim that charge packets self-localize. The research file's Gap 2 speculates about this, but nothing in the supplied material supports it, so we drop it. What does carry over is the structure of the dynamics.

For $N$ Kohn-Sham electrons with density $\rho(\mathbf{r},t)$ in a field $\mathbf{E}(\mathbf{r},t)$, the total electronic momentum $\mathbf{\Pi}$ obeys

$$ \frac{d\mathbf{\Pi}}{dt} = -\int \rho(\mathbf{r},t)\,\big[\mathbf{E}(\mathbf{r},t) + \nabla v_{\text{nuc}}(\mathbf{r})\big]\,d\mathbf{r} + \mathbf{C}[\gamma], $$

where $v_{\text{nuc}}$ is the potential of the fixed nuclei. The Hartree and exchange-correlation terms are absent because they exert zero net force. $\mathbf{C}[\gamma]$ is a correction computable from the density matrix $\gamma$ that appears only when the pseudopotentials are nonlocal, as in the ABACUS data. The dipole velocity is tied to $\mathbf{\Pi}$ by the continuity equation, again with a computable nonlocal correction.

The new element is architectural. A uniform phase factor $e^{i\boldsymbol{\kappa}\cdot\mathbf{r}}$ applied to every occupied orbital shifts $\mathbf{\Pi}$ by $N\boldsymbol{\kappa}$ and leaves the density untouched. This phase boost therefore acts as a clean "collective-coordinate knob," the discrete analog of the Melnikov et al. canonical transformation. At every rollout step the network proposes the next orbitals or density matrix. In parallel, $\mathbf{\Pi}$ is advanced by a symplectic (Verlet) step of the force law above, evaluated on the network's own predicted density. A phase boost then sets the proposal's momentum exactly to this value. The center-of-mass sector is thus integrated by physics and the internal sector by the network.

A modified KdV equation with self-consistent sources in non-uniform media and soliton dynamics motivates this split from the integrable side. It derives exact solitons driven by self-consistent sources and finds that the sources change velocity and trajectory while the internal balance preserves the shape. Driving acts mainly through collective coordinates, which is the behavior our architecture assumes.

A side benefit addresses a known pathology. Time-dependent density functional theory with ultrasoft pseudopotentials: Real-time electron propagation across a molecular junction propagates electron wave packets across a gold–benzene-dithiolate junction with a norm-conserving Crank–Nicolson scheme. It reports that its initial dipole "kick" slightly violates the f-sum rule. That rule says the total absorption strength of a molecule must equal its number of electrons, and it is equivalent to the initial momentum transfer being exactly $N\boldsymbol{\kappa}$. Our construction enforces that transfer by definition.

Research Program

The base propagator is Orbital Transformers for Predicting Wavefunctions in Time-Dependent Density Functional Theory (OrbEvo). It is an equivariant graph transformer that rolls out Kohn-Sham wavefunctions under a uniform field on about 5,000 QM9 molecules. It conditions on the field axis with SO(2) symmetry, because the field breaks full rotational symmetry, and it shows that messages built from the density matrix (OrbEvo-DM) beat pooled wavefunction features. We keep OrbEvo-DM unchanged and add the collective-momentum integrator and the phase-boost projection as a wrapper around it. The wrapper is cheap: the force integral is one quadrature over the predicted density per step, plus the nonlocal-pseudopotential commutator.

The first stage is a one-dimensional sanity model: a few Kohn-Sham electrons in a harmonic trap with soft-Coulomb interaction and a local exchange-correlation term. In this setting a uniform drive moves the density rigidly along the classical trajectory. This is an exact accelerating-solution analog of the forced soliton. A correct implementation must reproduce that trajectory to integrator precision whatever the network does.

The second stage is the OrbEvo QM9 setting itself.

The third stage tests two regimes where the constraint should matter most. One is longer horizons: the model is trained on the published 5-femtosecond windows and evaluated on newly generated 20-femtosecond trajectories for a held-out subset. The other is spatially varying fields with a linear gradient. Here OrbEvo's single-axis conditioning is no longer exact, yet the force law still holds, with the gradient entering through $\mathbf{E}(\mathbf{r},t)$.

Baselines

The do-nothing reference has two parts. The first predicts the ground-state density matrix for all times. The second predicts only the rigid response given by our exact collective integrator, with no internal dynamics. If the full model does not clearly beat the second, the network adds nothing beyond the physics.

The strongest existing method is OrbEvo-DM Orbital Transformers for Predicting Wavefunctions in Time-Dependent Density Functional Theory. It is the only learned real-time TDDFT propagator in the supplied material, and it already includes the right symmetry handling and the density-matrix inductive bias. For error, there is no reported OrbEvo wavefunction or dipole error on Lacuna. The file records only its cost comparison: rollouts take seconds per batch on a GPU, against 1.5 to 3.2 hours per molecule for ABACUS on a 24-core CPU. We will therefore retrain OrbEvo-DM ourselves on identical splits with the same compute and hyperparameter budget, and treat the retrained numbers as the reference. Beating it means a statistically clear reduction in induced-dipole error over the 20-femtosecond test window and in absorption-peak position error, with in-window error no worse than OrbEvo-DM.

The ground-truth reference is ABACUS real-time TDDFT itself. For one physical cross-check we compare the benzene π→π* peak, which Time-dependent density functional theory with ultrasoft pseudopotentials reports at 6.95 eV.

Three ablations isolate the key ingredient. The first, a soft-penalty model, is OrbEvo-DM trained with an added loss on the residual of the force law. The second, a learned-collective model, applies the phase boost but takes $\mathbf{\Pi}$ from a learned head instead of the exact force. The third, post-hoc correction, applies the boost only at inference to a network trained without it. All ablations share the backbone, data, and tuning budget.

Why not just…?

Why not just add the Ehrenfest relation as a soft loss term? This is the real threat, and it is ablation one. A penalty can make the residual small on average, but nothing stops errors from accumulating along the dipole direction during autoregressive rollout outside the training horizon. The hard projection makes the residual zero at every step. If the soft-penalty model matches our model on 20-femtosecond dipole drift and on the spectrum, we will report that the contribution is a loss term, not an architecture.

Why not just train OrbEvo-DM on more or longer trajectories? Data cost scales directly with ABACUS time, hours per molecule per 5 femtoseconds, so long-horizon data is exactly what is scarce. We will test this by giving the baseline twice the training molecules while our model keeps the original budget, and checking which effect is larger.

Why not just predict the dipole time series directly with a sequence model, since the spectrum is what is measured? A dipole-only model discards the density and orbitals. It cannot be evaluated against the self-consistent force law, and it cannot supply the density that downstream analyses use, such as transport or density-based surrogates like Graph neural networks for fast electron density estimation of molecules, liquids, and solids and Gaussian Plane-Wave Neural Operator for Electron Density Estimation. We will still train such a model as a check on spectrum metrics alone. If it wins there, that bounds what our approach adds for spectroscopy specifically.

Physics-informed neural networks and soliton-finding operators are not credible competitors at this scale. The nonlocal Hirota network in N-double poles solutions for nonlocal Hirota equation with nonzero boundary conditions using Riemann–Hilbert method and PINN algorithm reaches an L2 error of 2.69 × 10⁻⁴, but needs about 1,262 seconds per one-dimensional solution. The complex-valued operator of Complex-valued neural-operator-assisted soliton identification reaches 96–98% overlap on one-dimensional Gross–Pitaevskii solitons, but its authors flag the curse of dimensionality beyond one dimension. Neither applies to three-dimensional molecules.

Proposed Experiments

The one-dimensional trap test must show center-of-mass error at integrator precision for our model, while OrbEvo-style baselines drift. The QM9 test uses the OrbEvo split and measures three things: density-matrix error within the window, induced-dipole mean absolute error, and the error of the absorption spectrum obtained by Fourier transform, specifically peak positions in eV and the deviation of integrated oscillator strength from $N$. The long-horizon test trains on 5 femtoseconds and evaluates on 20 femtoseconds for a few hundred newly simulated held-out molecules, with error-growth curves for every model. The transfer test uses malonaldehyde configurations from MD17, as OrbEvo did. The final test uses field-gradient drives.

Two outcomes are decisive. The first is whether the exact collective integrator flattens the long-horizon dipole error curve compared with both OrbEvo-DM and the soft-penalty ablation. The second is whether the f-sum deviation falls to the integrator floor while the baselines show a measurable violation. If neither happens, the center-of-mass sector is not where the surrogates fail, and that negative result is itself worth reporting.