Phase-Faithful Surrogates for Real-Time TDDFT, Checked by Fano Lineshapes

Neural networks can now replace real-time time-dependent density functional theory (rt-TDDFT), the standard simulation of how a molecule's electrons respond to a laser pulse, at speeds hundreds to thousands of times faster than the original solver. These surrogates are judged almost entirely on quantities that ignore the quantum phase of the electrons, such as charges or dipole errors over a few femtoseconds. Yet the physics they are meant to reproduce depends on phase. The electrons' momentum and the interference between excitation pathways are both encoded in relative phases, and spectroscopy has long used Fano lineshapes to read those phases out. A Fano lineshape is the lopsided absorption peak that appears when a sharp resonance interferes with a broad background. Its degree of asymmetry is set by the relative phase of the two pathways. We propose to use that asymmetry as the test of whether a surrogate keeps phase correct over long rollouts. We also propose a cheap per-step correction that enforces exact phase-dependent identities of TDDFT dynamics. When those identities cannot be met without overriding the network, the correction relaxes them gracefully, using the "optimal-decay" construction from safety-critical control, rather than failing. The amount of relaxation it needs then tells us when to hand control back to the expensive solver for a few steps.

The two-sentence version is this. Current rt-TDDFT surrogates are evaluated over windows too short to see interference, and nothing in them stops relative phase from drifting. We add a bounded, always-solvable physics correction whose relaxation signal triggers brief fallbacks to the exact solver, and we measure success by whether long-rollout spectra recover the correct Fano asymmetry at a given compute budget.

Why this gap is real

Three lines of work replace rt-TDDFT with learned models, and each one runs into phase. Machine Learning Models Capture Plasmon Dynamics in Ag Nanoparticles uses a recurrent network on histories of atomic charges to predict silver nanoparticle plasmon dynamics. It reports a 200× speedup for Ag55 and a 4000× speedup for Ag561, with qualitative accuracy up to about 32 fs. The authors state that scalar charges discard phase and so cannot resolve hot-carrier populations. Orbital Transformers for Predicting Wavefunctions in Time-Dependent Density Functional Theory (OrbEvo) predicts orbital coefficients under a field. It removes the trivial global phase rotation and builds the density matrix explicitly. Its density-matrix variant beats wavefunction pooling on dipoles and spectra, but it is evaluated only over 5 fs rollouts. Learning Potentials of Quantum Systems using Deep Neural Networks recovers static potentials from densities and says explicitly that the time-dependent case needs the missing phase. These papers motivate the proposal: each documents phase as the open weakness.

The 5 fs window matters more than it seems. A spectrum computed from a dipole trace of length $T$ cannot resolve features narrower than roughly $h/T$, which is the standard time–frequency resolution limit. With $h \approx 4.14$ eV·fs, a 5 fs rollout gives about 0.8 eV resolution, far too coarse to see whether a resonance is lopsided. Phase errors that would distort a lineshape are therefore invisible in current evaluations by construction. Seeing them requires rollouts of 50 fs or more, which is where surrogates become unstable.

The measurement principle comes from quantum optics. Detecting the relative phase between different frequency components of a photon using a three-level Λ atom coupled to a waveguide shows that the relative phase between scattering and non-scattering pathways appears directly as transmission-lineshape asymmetry, with no reference beam. Phase Modulation in Rydberg Dressed Multi-Wave Mixing processes maps an interaction-induced nonlinear phase shift onto spectral asymmetry in experiment. We borrow their principle, not their systems. A computed absorption peak near a resonance at frequency $\omega_r$ with width $\Gamma$ is fit by

$$\sigma(\omega) \propto \frac{(q+\epsilon)^2}{1+\epsilon^2}, \qquad \epsilon = \frac{2(\omega-\omega_r)}{\Gamma},$$

where $q$ is the Fano asymmetry parameter. The quantity $1/q$ shrinks to zero for a symmetric peak and grows as the pathway phases depart from that limit. A surrogate that places $\omega_r$ correctly but gets $q$ wrong has learned peak energies but not phase. Measuring quantum states: Experimental setup for measuring the spatial density matrix supplies the matching static fact: the off-diagonal phase elements of the density matrix carry momentum, and a diagonal density alone discards them.

Core Claim

Our first claim concerns the failure. Surrogates that are accurate on short-window dipoles will drift in relative phase at longer times. This will show up first as wrong Fano asymmetry, while peak positions stay roughly correct.

Our second claim concerns the remedy, which is a correction that targets phase specifically. Making the density matrix physically valid is not enough. Orthonormalizing predicted orbitals already guarantees a Hermitian, trace-preserving, positive and idempotent density matrix, and it says nothing about relative phase. The identities that constrain phase are dynamical, and TDDFT satisfies two of them along the field direction.

The first is energy balance. The change in the system's energy over a step must equal the work done by the field on the dipole.

The second is dipole–momentum consistency. The rate of change of the dipole must equal the momentum expectation, which depends on the imaginary, phase-carrying part of the density matrix. This identity is exact for local potentials. A finite basis and nonlocal pseudopotentials break it slightly, so we set its tolerance from the residual of the reference solver rather than assuming zero.

At each step we apply a small correction $\delta$ to the predicted orbital coefficients $c_t$:

$$\min_{\delta,\ \omega \ge 0} \ \|\delta\|^2 + \lambda(\omega-1)^2 \quad \text{s.t.}\quad |r_k(c_t+\delta)| \le \omega\,\tau_k \ \ (k=1,2), \qquad \|\delta\|_\infty \le \Delta.$$

Here $r_1$ and $r_2$ are the residuals of the two identities, linearized around the prediction, and $\tau_k$ are their tolerances measured on the reference solver. The bound $\Delta$ limits how far the correction may override the network. The variable $\omega$ is a relaxation factor with nominal value 1, and $\lambda$ sets how strongly it is pulled back to that value. Orthonormalization is applied afterward.

This problem mirrors the structure of a control safety filter. The correction plays the role of the control input, the bound $\Delta$ plays the role of actuator limits, and the two identities play the role of the barrier constraints. As in Control Barrier Functions: Theory and Applications, the filter becomes infeasible when bounded inputs cannot satisfy the constraints. We reuse the remedy of Safety-Critical Control using Optimal-decay Control Barrier Function with Guaranteed Point-wise Feasibility, which introduces a decay variable that loosens the barrier just enough to keep the program solvable. That paper showed a standard filter freezing on an adaptive-cruise benchmark at 32 m/s where the relaxed version braked safely. On the Properties of Optimal-Decay Control Barrier Functions derives closed-form solutions. Our program has only a handful of scalar constraints, so it can be solved in closed form or nearly so, and its cost per step is negligible next to the network.

One caveat from that work also applies here. Satisfying the constraints at each step does not guarantee they stay satisfiable later, as Zeng et al. note. We therefore treat a rising $\omega$ as evidence that the surrogate has left reliable territory. When $\omega$ stays above a threshold $\omega^\ast$, we run rt-TDDFT for a short segment and restart the surrogate from that state. Sweeping $\omega^\ast$ gives an explicit cost–accuracy curve.

Research Program

We will first generate long reference trajectories, because none in the file exceed about 32 fs. We will extend the OrbEvo protocol on its QM9 subset and on the MD17 malondialdehyde configurations to 50–100 fs under weak δ-kick and pulsed fields. From these runs we will keep only systems whose reference spectra contain peaks that are clearly asymmetric, using a Fano fit with a confidence interval on $1/q$ that excludes zero. This filter matters because finite-basis molecular spectra may contain few genuine interference features. If too few survive, the claim weakens to the subset that does.

On the Ag55–Ag147 nanoparticles we will test whether the reference plasmon peak is asymmetric where it overlaps interband transitions. We include that system only if the reference shows it.

To calibrate the metric, we will build a one-dimensional model of a bound state coupled to a discretized quasi-continuum, in the style of the synthetic benchmarks of Sehanobish et al. There we can dial the coupling phase directly and confirm that the fitted $q$ recovers it.

We will then train OrbEvo-DM-style surrogates. We take OrbEvo's architecture as the backbone and its delta transformation unchanged, and add the correction and fallback at inference only. A second variant also backpropagates through the correction during training, to test whether in-loop constraints help beyond inference-time repair.

The analysis has three questions. Where in time do $q$ errors begin, compared with dipole errors? Does $\omega$ predict true error before it becomes visible? What fraction of steps must fall back to the solver to reach a target $q$ accuracy?

Baselines

The naive reference point is the unconstrained surrogate rolled out to the full horizon: OrbEvo-DM with its published architecture, retrained on our longer trajectories. It defines the error the proposal must reduce. The cost ceiling and accuracy reference is full rt-TDDFT. OrbEvo reports that rt-TDDFT takes 1.5 to 3.2 hours per molecule on a 24-core CPU, against seconds per batch on a single GPU for the surrogate.

The strongest existing method is OrbEvo-DM itself. No reported result was found on Lacuna for OrbEvo's dipole or spectrum error values, only its speed figures and the qualitative finding that the density-matrix variant beats wavefunction pooling. We will therefore reproduce it rather than compare against a quoted number. For the nanoparticles, the charge-history network of Habib et al. is the comparison, with its reported 200× (Ag55) and 4000× (Ag561) speedups and roughly 32 fs qualitative horizon as the reference point.

A cost-matched classical competitor is cheaper rt-TDDFT, using a larger time step or a smaller basis, tuned to match our hybrid's wall-clock time. If it reaches the same $q$ accuracy at the same cost, the surrogate route is not justified for this question.

Three ablations isolate the key ingredient. The first is orthonormalization alone, which gives physical validity without the dynamical identities. The second enforces the identities as hard constraints with no relaxation variable, and rejects the step and calls the solver whenever the program is infeasible. This is the analogue of the accept-reject safeguard in Derivative Informed Learning of Exchange-Correlation Functionals, which rejects training updates that destabilize the SCF solver. The third uses a fixed quadratic penalty with a hand-tuned weight in place of $\omega$.

For the fallback policy, we compare $\omega$-triggered fallback against fallback on a uniform schedule and against an ensemble-disagreement trigger in the spirit of Less is more: Sampling chemical space with active learning, at equal numbers of solver steps.

All variants share training data, backbone, rollout horizon, and a fixed hyperparameter search budget. The proposal beats a baseline if it achieves lower error in $1/q$ at matched wall-clock time, or matched error at lower cost, across held-out molecules. Peak position and dipole accuracy must not degrade relative to OrbEvo-DM.

Why not just…?

Why not just orthonormalize the predicted orbitals each step? This is the most serious threat, because it is nearly free and enforces every static physical constraint exactly. Our argument is that static validity places no constraint on relative phase between orbitals, so lineshapes can still drift. That argument is a hypothesis, not evidence. The orthonormalize-only ablation settles it directly. If it matches the full method on $q$ error, we will report that the dynamical identities are unnecessary. The contribution would then reduce to the long-horizon Fano benchmark and the cost-accuracy analysis of fallback.

Why not just fine-tune on longer rollouts with a spectrum loss? Training directly on spectra is a strong alternative, and it may reduce $q$ error on molecules seen in training. Two things favor the correction instead. It needs no extra trajectories at training time. It also gives a per-step relaxation signal that a fine-tuned model lacks, and that signal is what drives the cost–accuracy trade. We will include a spectrum-loss fine-tuned OrbEvo-DM and test both on held-out molecules. The proposal loses if the fine-tuned model matches it out of distribution without any fallback.

Why not just run cheaper classical rt-TDDFT? The surrogate speed figures already reported suggest large gains, but those compare against full-accuracy runs, not degraded ones. A coarser time step or basis may preserve lineshapes well enough. The cost-matched baseline above answers this directly.

Why not use a fixed penalty instead of the optimal-decay variable? Zeng et al. showed fixed-parameter filters failing outright when constraints become infeasible. Resilient Control: Compromising to Adapt reports sudden "cliff-edge" failures for fixed-penalty designs, where dynamic slack degrades gracefully. A fixed penalty also yields no interpretable trigger for fallback. Whether these control results carry over to electron dynamics is exactly what the penalty ablation tests.

Decisive Checks

The proposal stands or falls on four observations. First, on the reference-filtered systems, unconstrained OrbEvo-DM must show a growing $1/q$ error by 50 fs even while peak positions remain accurate. If it does not, the phase-drift premise is false. Second, the full method must beat the orthonormalize-only ablation on $q$. Third, $\omega$-triggered fallback must dominate uniform and ensemble-disagreement triggers on the cost–accuracy curve. Fourth, on the one-dimensional calibration model, the fitted $q$ must track the imposed pathway phase. Without that, the metric cannot be trusted as a phase readout.