arxiv: 2604.25137 · v1 · submitted 2026-04-28 · 🪐 quant-ph · cs.LG· physics.chem-ph· physics.comp-ph

Recognition: unknown

Quantum Dynamics via Score Matching on Bohmian Trajectories

Lei Wang

Authors on Pith no claims yet

Pith reviewed 2026-05-07 16:56 UTC · model grok-4.3

classification 🪐 quant-ph cs.LGphysics.chem-phphysics.comp-ph

keywords quantum dynamicsscore matchingBohmian trajectoriesnormalizing flowsSchrödinger equationneural networkswave-packet dynamicsgenerative modeling

0 comments

The pith

Minimizing a self-consistent score-matching loss on Bohmian trajectories recovers exact Schrödinger dynamics for nodeless wave functions.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes a method to evolve quantum states by training a neural network to learn the score function along deterministic Bohmian particle paths. These paths move under a velocity field that depends on the score of the evolving probability density, turning the dynamics into a continuous normalizing flow. A self-consistent objective is minimized that penalizes the mismatch between the network's score and the score implied by the density produced by the flow itself. When this loss reaches zero, the resulting evolution satisfies the time-dependent Schrödinger equation provided the wave function remains nodeless. This condition holds for many vibrational problems in atoms and molecules, allowing the approach to be demonstrated on wave-packet splitting and anharmonic chain vibrations.

Core claim

Parametrizing the score with a neural network and minimizing the self-consistent Fisher divergence between the network output and the score of the density evolved under the Bohmian velocity field derived from that score yields a zero-loss solution that satisfies both the continuity equation and the quantum Hamilton-Jacobi equation. Consequently the learned dynamics recover the full time-dependent Schrödinger equation for nodeless wave functions. The non-crossing nature of the trajectories guarantees that the mapping remains a valid continuous normalizing flow.

What carries the argument

Self-consistent Fisher divergence minimization between a neural-network score and the score of the probability density induced by the Bohmian flow whose velocity is itself determined by that score.

If this is right

Wave-packet splitting in a double-well potential can be simulated by evolving only an ensemble of trajectories guided by the learned score.
Anharmonic vibrations of a Morse chain can be computed without discretizing the full many-dimensional wave function.
Zero loss guarantees exact recovery of Schrödinger evolution whenever the wave function remains nodeless.
Time-dependent quantum problems are recast as training tasks in score-based generative modeling.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same self-consistent objective could be paired with more advanced normalizing-flow architectures to improve sampling efficiency in higher-dimensional systems.
Because the method is demonstrated on atomic vibrations, it suggests a route to computing vibrational spectra directly from trajectory ensembles rather than from grid-based wave functions.

Load-bearing premise

The minimization procedure is assumed to reach the true score globally without becoming trapped by approximation errors or instabilities, and the target wave functions must stay nodeless so that trajectories never cross.

What would settle it

Apply the trained network to the exactly solvable harmonic oscillator, extract the time-evolved density from the trajectories, and check whether it matches the known analytic Gaussian wave-packet solution to numerical precision at multiple times.

Figures

Figures reproduced from arXiv: 2604.25137 by Lei Wang.

**Figure 1.** Figure 1: FIG. 1. Quantum dynamics as a continuous normalizing flow. view at source ↗

**Figure 3.** Figure 3: shows the training convergence over 45,000 epochs. The occasional spikes visible in panels (a) and (b) are caustic events: |det F| transiently becomes small, so the score target F −T[∇x(0) ln ρ0 − ∇x(0) ln |det F|] in Eq. (7) inflates and the network takes an outsized gradient step. The system self-heals within tens of epochs: the learned Q strengthens, |det F| recovers, and both the loss and energy erro… view at source ↗

**Figure 5.** Figure 5: FIG. 5. Learned score field view at source ↗

read the original abstract

We solve the time-dependent Schr\"odinger equation by learning the score function, the gradient of the log-probability density, on Bohmian trajectories. In Bohm's formulation of quantum mechanics, particles follow deterministic paths under the classical potential supplemented by a quantum potential depending on the score function of the evolving density. These non-crossing Bohmian trajectories form a continuous normalizing flow governed by the score. We parametrize the score with a neural network and minimize a self-consistent Fisher divergence between the network and the score of the resulting density. We prove that the zero-loss minimizer of this self-consistent objective recovers Schr\"odinger dynamics for nodeless wave functions, a condition naturally met in quantum vibrations of atoms. We demonstrate the approach on wavepacket splitting in a double-well potential and anharmonic vibrations of a Morse chain. By recasting real-time quantum dynamics as a self-consistent score-driven normalizing flow, this framework opens the time-dependent Schr\"odinger equation to the rapidly advancing toolkit of modern generative modeling.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper turns TDSE simulation into self-consistent neural score matching on Bohmian trajectories and proves recovery of exact dynamics for nodeless wavefunctions.

read the letter

The main point is that this work recasts real-time quantum evolution as a generative modeling task: a neural network learns the score on the fly from trajectories whose velocity depends on that same score. They show a proof that zero loss on the self-consistent Fisher divergence recovers Schrödinger dynamics exactly when the wavefunction has no nodes, which covers many vibrational problems. The demos on double-well splitting and Morse oscillator chains are straightforward and run without reported instabilities on those cases.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes solving the time-dependent Schrödinger equation by parametrizing the score function (gradient of log-density) with a neural network and minimizing a self-consistent Fisher divergence objective on Bohmian trajectories. These trajectories form a continuous normalizing flow whose velocity depends on the learned score. The central theoretical claim is a proof that any zero-loss minimizer of this objective recovers exact Schrödinger dynamics for nodeless wave functions. The approach is demonstrated numerically on wave-packet splitting in a double-well potential and anharmonic vibrations of a Morse chain.

Significance. If the zero-loss recovery result holds and the optimization reaches it in practice, the work would usefully recast real-time quantum dynamics as a score-based generative modeling task, potentially allowing the TDSE to benefit from advances in continuous normalizing flows and score matching. The explicit proof for the nodeless case (relevant to many vibrational problems) and the connection to Bohmian mechanics are concrete strengths that distinguish the contribution from purely heuristic ML-for-quantum-dynamics papers.

major comments (2)

[theoretical derivation of zero-loss minimizer] Proof of zero-loss recovery (theoretical section): the argument correctly shows that an exact zero of the self-consistent Fisher divergence implies the TDSE under the nodeless assumption, but does not establish uniqueness of the fixed point or rule out other self-consistent solutions that satisfy the objective yet fail to reproduce Schrödinger evolution. Because trajectories are generated on-the-fly from the current network, this leaves open whether the optimization landscape contains non-physical attractors.
[demonstrations on double-well and Morse potentials] Numerical experiments (double-well and Morse-chain sections): the demonstrations are qualitative only; no quantitative error metrics (e.g., L2 deviation from exact TDSE solutions, conservation of energy or norm, or convergence of the loss to machine zero) are reported, nor is there analysis of discretization artifacts or batch-size effects on the self-consistent loop. This makes it impossible to assess whether practical training reaches the regime where the proof applies.

minor comments (2)

[methods] Notation for the self-consistent objective and the Fisher divergence should be introduced with an explicit equation number early in the methods section to improve readability.
[discussion] The manuscript would benefit from a short discussion of how the nodeless condition is verified or enforced for the target systems, even if it holds for the chosen examples.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the positive assessment of our work's significance and for the constructive comments. We address each major point below and describe the revisions we will undertake.

read point-by-point responses

Referee: Proof of zero-loss recovery (theoretical section): the argument correctly shows that an exact zero of the self-consistent Fisher divergence implies the TDSE under the nodeless assumption, but does not establish uniqueness of the fixed point or rule out other self-consistent solutions that satisfy the objective yet fail to reproduce Schrödinger evolution. Because trajectories are generated on-the-fly from the current network, this leaves open whether the optimization landscape contains non-physical attractors.

Authors: We appreciate the referee's precise reading. Our derivation shows that any exact zero of the self-consistent objective recovers the TDSE for nodeless wave functions, as stated. We agree that the proof does not establish uniqueness of this fixed point or exclude other self-consistent solutions, and that on-the-fly trajectory generation leaves open the possibility of non-physical attractors. In the revised manuscript we will add an explicit discussion of this limitation in the theoretical section, noting that the nodeless condition together with the continuity of the Bohmian flow strongly constrains the solution space while acknowledging that practical optimization may require suitable initialization to reach the physical attractor. We present this as an important open question rather than a fully settled claim. revision: partial
Referee: Numerical experiments (double-well and Morse-chain sections): the demonstrations are qualitative only; no quantitative error metrics (e.g., L2 deviation from exact TDSE solutions, conservation of energy or norm, or convergence of the loss to machine zero) are reported, nor is there analysis of discretization artifacts or batch-size effects on the self-consistent loop. This makes it impossible to assess whether practical training reaches the regime where the proof applies.

Authors: We agree that the current numerical demonstrations are qualitative. In the revised manuscript we will add quantitative metrics, including L2 deviations from reference TDSE solutions for the double-well example, time series of energy and norm conservation, and plots of loss convergence during self-consistent training. We will also include a brief analysis of discretization step size and batch-size effects on the stability of the self-consistent loop to demonstrate that the reported trajectories operate in the regime where the zero-loss guarantee is expected to hold. revision: yes

Circularity Check

0 steps flagged

No significant circularity in the derivation chain

full rationale

The paper defines a self-consistent Fisher divergence objective whose fixed point is the score matching the density induced by trajectories driven by that score. It then supplies an explicit proof that any zero-loss solution of this objective recovers the time-dependent Schrödinger equation for nodeless wave functions. Because the proof supplies an independent mathematical reduction from the fixed-point condition to the TDSE (rather than the equivalence being true by definition or by renaming), the central claim does not collapse to its inputs. No load-bearing self-citations, fitted parameters presented as predictions, or smuggled ansatzes appear in the derivation. The self-consistency is a deliberate fixed-point formulation whose correctness is verified externally by the proof, making the overall argument self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claim rests on standard concepts from Bohmian mechanics and score matching, with the self-consistency and nodeless proof as the main additions. No new physical entities are postulated.

free parameters (1)

Neural network weights
The score function is parametrized by a neural network whose parameters are optimized to minimize the self-consistent Fisher divergence.

axioms (2)

domain assumption Bohmian trajectories form a continuous normalizing flow governed by the score function
Invoked in the abstract to link the deterministic paths to the density evolution.
domain assumption Zero loss of the self-consistent objective recovers exact Schrödinger dynamics for nodeless wave functions
The key theoretical result claimed in the abstract.

pith-pipeline@v0.9.0 · 5474 in / 1463 out tokens · 53850 ms · 2026-05-07T16:56:12.748861+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

45 extracted references · 11 canonical work pages · 5 internal anchors

[1]

Meyer, U

H.-D. Meyer, U. Manthe, and L. S. Cederbaum, The multi-configurational time-dependent Hartree approach, Chem. Phys. Lett.165, 73 (1990)

1990
[2]

Carleo and M

G. Carleo and M. Troyer, Solving the quantum many- body problem with artificial neural networks, Science 355, 602 (2017)

2017
[3]

C. L. Lopreore and R. E. Wyatt, Quantum wave packet dynamics with trajectories, Physical Review Letters82, 5190 (1999)

1999
[4]

Sales Mayor, A

F. Sales Mayor, A. Askar, and H. A. Rabitz, Quantum fluid dynamics in the Lagrangian representation and ap- plications to photodissociation problems, The Journal of Chemical Physics111, 2423 (1999)

1999
[5]

R. E. Wyatt,Quantum Dynamics with Trajectories: In- troduction to Quantum Hydrodynamics, Interdisciplinary Applied Mathematics, Vol. 28 (Springer, 2005)

2005
[6]

Madelung, Quantentheorie in hydrodynamischer Form, Zeitschrift f¨ ur Physik40, 322 (1927)

E. Madelung, Quantentheorie in hydrodynamischer Form, Zeitschrift f¨ ur Physik40, 322 (1927)

1927
[7]

de Broglie, La m´ ecanique ondulatoire et la structure atomique de la mati` ere et du rayonnement, Journal de Physique et le Radium8, 225 (1927)

L. de Broglie, La m´ ecanique ondulatoire et la structure atomique de la mati` ere et du rayonnement, Journal de Physique et le Radium8, 225 (1927)

1927
[8]

Bohm, A suggested interpretation of the quantum the- ory in terms of “hidden” variables

D. Bohm, A suggested interpretation of the quantum the- ory in terms of “hidden” variables. I, Physical Review85, 166 (1952)

1952
[9]

Bohm, A suggested interpretation of the quantum the- ory in terms of “hidden” variables

D. Bohm, A suggested interpretation of the quantum the- ory in terms of “hidden” variables. II, Physical Review 85, 180 (1952)

1952
[10]

R. T. Q. Chen, Y. Rubanova, J. Bettencourt, and D. Du- venaud, Neural ordinary differential equations, Advances in Neural Information Processing Systems31(2018)

2018
[11]

Zhang, W

L. Zhang, W. E, and L. Wang, Monge-Amp` ere flow for generative modeling, arXiv:1809.10188 (2018)

work page arXiv 2018
[12]

Song and S

Y. Song and S. Ermon, Generative modeling by estimat- ing gradients of the data distribution, inAdvances in Neural Information Processing Systems, Vol. 32 (2019)

2019
[13]

Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, Score-based generative mod- eling through stochastic differential equations, inInter- national Conference on Learning Representations(2021) arXiv:2011.13456

work page internal anchor Pith review arXiv 2021
[14]

P. R. Holland,The Quantum Theory of Motion: An Account of the de Broglie–Bohm Causal Interpretation of Quantum Mechanics(Cambridge University Press, 1993)

1993
[15]

Garashchuk and V

S. Garashchuk and V. A. Rassolov, Quantum dynamics with bohmian trajectories: energy conserving approxi- mation to the quantum potential, Chemical Physics Let- ters376, 358 (2003)

2003
[16]

Garashchuk and V

S. Garashchuk and V. A. Rassolov, Energy conserving ap- proximations to the quantum potential: Dynamics with linearized quantum force, Journal of Chemical Physics 120, 1181 (2004)

2004
[17]

Hyv¨ arinen, Estimation of non-normalized statistical models by score matching, Journal of Machine Learning Research6, 695 (2005)

A. Hyv¨ arinen, Estimation of non-normalized statistical models by score matching, Journal of Machine Learning Research6, 695 (2005)

2005
[18]

J. B. Maddox and E. R. Bittner, Estimating Bohm’s quantum force using Bayesian statistics, The Journal of Chemical Physics119, 6465 (2003)

2003
[19]

T. Chen, B. Xu, C. Zhang, and C. Guestrin, Training deep nets with sublinear memory cost, arXiv:1604.06174 (2016)

work page internal anchor Pith review arXiv 2016
[20]

Perez, F

E. Perez, F. Strub, H. de Vries, V. Dumoulin, and A. Courville, FiLM: Visual reasoning with a general con- ditioning layer, inProceedings of the AAAI Conference on Artificial Intelligence, Vol. 32 (2018)

2018
[21]

F´ enyes, Eine wahrscheinlichkeitstheoretische Begr¨ undung und Interpretation der Quantenmechanik, Zeitschrift f¨ ur Physik132, 81 (1952)

I. F´ enyes, Eine wahrscheinlichkeitstheoretische Begr¨ undung und Interpretation der Quantenmechanik, Zeitschrift f¨ ur Physik132, 81 (1952)

1952
[22]

Nelson, Derivation of the Schr¨ odinger equation from Newtonian mechanics, Physical Review150, 1079 (1966)

E. Nelson, Derivation of the Schr¨ odinger equation from Newtonian mechanics, Physical Review150, 1079 (1966)

1966
[23]

Maoutsa, S

D. Maoutsa, S. Reich, and M. Opper, Interacting particle solutions of Fokker–Planck equations through gradient– log–density estimation, Entropy22, 802 (2020)

2020
[24]

Z. Shen, Z. Wang, S. Kale, A. Ribeiro, A. Karbasi, and H. Hassani, Self-consistency of the Fokker–Planck equa- tion, inProceedings of the 35th Conference on Learning Theory, Proceedings of Machine Learning Research, Vol. 178 (2022) pp. 817–841

2022
[25]

N. M. Boffi and E. Vanden-Eijnden, Probability flow so- lution of the Fokker–Planck equation, Machine Learning: Science and Technology4, 035012 (2023)

2023
[26]

L. Li, S. Hurault, and J. Solomon, Self-consistent veloc- ity matching of probability flows, inAdvances in Neural Information Processing Systems, Vol. 36 (2023)

2023
[27]

Wu and Q

X. Wu and Q. Liao, Adaptive probability flow resid- ual minimization for high-dimensional Fokker–Planck equations, arXiv preprint arXiv:2512.19196 (2025), arXiv:2512.19196

work page arXiv 2025
[28]

Flow Matching for Generative Modeling

Y. Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le, Flow matching for generative modeling, inInter- national Conference on Learning Representations(2023) arXiv:2210.02747

work page internal anchor Pith review arXiv 2023
[29]

X. Liu, C. Gong, and Q. Liu, Flow straight and fast: Learning to generate and transfer data with rectified flow, inInternational Conference on Learning Representations (2023) arXiv:2209.03003

work page internal anchor Pith review arXiv 2023
[30]

M. S. Albergo and E. Vanden-Eijnden, Building nor- malizing flows with stochastic interpolants, inInterna- tional Conference on Learning Representations(2023) arXiv:2209.15571

work page internal anchor Pith review arXiv 2023
[31]

Orlova, A

E. Orlova, A. Ustimenko, R. Jiang, P. Y. Lu, and R. Wil- lett, Deep stochastic mechanics, inProceedings of the 41st International Conference on Machine Learning, Proceed- ings of Machine Learning Research, Vol. 235 (PMLR,
[32]

38779–38814, arXiv:2305.19685

pp. 38779–38814, arXiv:2305.19685

work page arXiv
[33]

Carleo, F

G. Carleo, F. Becca, M. Schir´ o, and M. Fabrizio, Lo- calization and glassy dynamics of many-body quantum systems, Sci. Rep.2, 243 (2012)

2012
[34]

P. A. M. Dirac, Note on exchange phenomena in the Thomas atom, Math. Proc. Cambridge Philos. Soc.26, 376 (1930)

1930
[35]

A. D. McLachlan, A variational solution of the time- dependent Schr¨ odinger equation, Mol. Phys.8, 39 (1964)

1964
[36]

Raissi, P

M. Raissi, P. Perdikaris, and G. E. Karniadakis, Physics- informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations, Journal of Computational Physics378, 686 (2019)

2019
[37]

E. Hou, Y. Liu, L. Zhang, D. Ye, L. Wang, and H. Wang, A global spacetime optimization approach to the real-space time-dependent Schr¨ odinger equation, arXiv preprint arXiv:2511.12983 (2026), arXiv:2511.12983 [quant-ph]

work page arXiv 2026
[38]

Sinibaldi, D

A. Sinibaldi, D. Hendry, F. Vicentini, and G. Carleo, 7 Time-dependent neural galerkin method for quantum dy- namics, Phys. Rev. Lett.136, 120402 (2026)

2026
[39]

H. Wu, S. Liu, X. Ye, and H. Zhou, A parameterized Wasserstein Hamiltonian flow approach for solving the Schr¨ odinger equation, arXiv:2505.11762 (2025)

work page arXiv 2025
[40]

D. M. Ceperley, Fermion nodes, J. Stat. Phys.63, 1237 (1991)

1991
[41]

T. C. Wallstrom, Inequivalence between the Schr¨ odinger equation and the Madelung hydrodynamic equations, Physical Review A49, 1613 (1994)

1994
[42]

H. Xie, L. Zhang, and L. Wang, Ab-initio study of inter- acting fermions at finite temperature with neural canon- ical transformation, Journal of Machine Learning1, 38 (2022)

2022
[43]

H. Xie, L. Zhang, and L. Wang,m ∗ of two-dimensional electron gas: A neural canonical transformation study, SciPost Phys.14, 154 (2023)

2023
[44]

Z. Li, H. Xie, X. Dong, and L. Wang, Deep variational free energy calculation of hydrogen hugoniot, Phys. Rev. Lett.136, 076504 (2026)

2026
[45]

FFJORD: Free-form Continuous Dynamics for Scalable Reversible Generative Models

W. Grathwohl, R. T. Q. Chen, J. Bettencourt, I. Sutskever, and D. Duvenaud, FFJORD: Free-form con- tinuous dynamics for scalable reversible generative mod- els, inInternational Conference on Learning Representa- tions(2019) arXiv:1810.01367. 8 End Matter Proof of the exactness theorem We prove the Exactness Theorem by constructing the wave function Ψ = √ρ...

work page Pith review arXiv 2019