Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Faster Molecular Dynamics with Neural Network Potentials via Distilled Multiple Time-Stepping and Non-Conservative Forces

Nicolaï Gouraud, Côme Cattin, Thomas Plé, Olivier Adjoua, Louis Lagardère, Jean-Philip Piquemal

TL;DR

This work addresses the challenge of accelerating molecular dynamics with neural network potentials without sacrificing near-ab initio accuracy. It introduces DMTS-NC, a distilled non-conservative force model used in a dual-level RESPA-based MTS integrator, augmented by Hydrogen Mass Repartitioning and a runtime revert mechanism to maintain stability. The NC model is trained by distillation from a large FeNNix-Bio1(M) model, achieving an MAE of $1.46$ kcal/mol/Å and RMSE of $2.33$ kcal/mol/Å, and enabling larger external time steps while preserving sampling quality. Across bulk water and solvated proteins, DMTS-NC delivers substantial speedups (up to $\sim 431\%$ over STS and $\sim 31-16\%$ over conservative/fine-tuned DMTS) with minimal bias in structural and thermodynamic properties, illustrating a practical path toward deploying NN potentials in large-scale MD simulations.

Abstract

Following our previous work (J. Phys. Chem. Lett., 2026, 17, 5, 1288-1295), we propose the DMTS-NC approach, a distilled multi-time-step (DMTS) strategy using non conservative (NC) forces to further accelerate atomistic molecular dynamics simulations using foundation neural network models. There, a dual-level reversible reference system propagator algorithm (RESPA) formalism couples a target accurate conservative potential to a simplified distilled representation optimized for the production of non-conservative forces. Despite being non-conservative, the distilled architecture is designed to enforce key physical priors, such as equivariance under rotation and cancellation of atomic force components. These choices facilitate the distillation process and therefore improve drastically the robustness of simulation, significantly limiting the "holes" in the simpler potential, thus achieving excellent agreement with the forces data. Overall, the DMTS-NC scheme is found to be more stable and efficient than its conservative counterpart with additional speedups reaching 15-30% over DMTS. Requiring no finetuning steps, it is easier to implement and can be pushed to the limit of the systems physical resonances to maintain accuracy while providing maximum efficiency. As for DMTS, DMTS-NC is applicable to any neural network potential.

Faster Molecular Dynamics with Neural Network Potentials via Distilled Multiple Time-Stepping and Non-Conservative Forces

TL;DR

This work addresses the challenge of accelerating molecular dynamics with neural network potentials without sacrificing near-ab initio accuracy. It introduces DMTS-NC, a distilled non-conservative force model used in a dual-level RESPA-based MTS integrator, augmented by Hydrogen Mass Repartitioning and a runtime revert mechanism to maintain stability. The NC model is trained by distillation from a large FeNNix-Bio1(M) model, achieving an MAE of kcal/mol/Å and RMSE of kcal/mol/Å, and enabling larger external time steps while preserving sampling quality. Across bulk water and solvated proteins, DMTS-NC delivers substantial speedups (up to over STS and over conservative/fine-tuned DMTS) with minimal bias in structural and thermodynamic properties, illustrating a practical path toward deploying NN potentials in large-scale MD simulations.

Abstract

Following our previous work (J. Phys. Chem. Lett., 2026, 17, 5, 1288-1295), we propose the DMTS-NC approach, a distilled multi-time-step (DMTS) strategy using non conservative (NC) forces to further accelerate atomistic molecular dynamics simulations using foundation neural network models. There, a dual-level reversible reference system propagator algorithm (RESPA) formalism couples a target accurate conservative potential to a simplified distilled representation optimized for the production of non-conservative forces. Despite being non-conservative, the distilled architecture is designed to enforce key physical priors, such as equivariance under rotation and cancellation of atomic force components. These choices facilitate the distillation process and therefore improve drastically the robustness of simulation, significantly limiting the "holes" in the simpler potential, thus achieving excellent agreement with the forces data. Overall, the DMTS-NC scheme is found to be more stable and efficient than its conservative counterpart with additional speedups reaching 15-30% over DMTS. Requiring no finetuning steps, it is easier to implement and can be pushed to the limit of the systems physical resonances to maintain accuracy while providing maximum efficiency. As for DMTS, DMTS-NC is applicable to any neural network potential.
Paper Structure (10 sections, 8 equations, 2 figures, 3 tables, 1 algorithm)

This paper contains 10 sections, 8 equations, 2 figures, 3 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Methods
  3. Model architecture.
  4. Model training.
  5. MTS integrator.
  6. Hydrogen Mass Repartitioning.
  7. Preventing rare disagreements.
  8. Numerical results
  9. Bulk water
  10. Solvated proteins

Figures (2)

  • Figure 1: Sampling properties of bulk water simulations. All data was collected from 2 nanosecond simulations, with one frame saved per picosecond. As the graphs above show, radial distributions are not affected by the use of DMTS-NC compared to DMTS or STS, temperatures fit the equilibrium theoretical distribution, and samples of potential energies, especially represented on the quantile-quantile plot, highlight the low-bias constant-variance deviation of the distributions from STS to DMTS-NC, as expected for a multi-time-step method in a stable regime.
  • Figure 2: Time evolution of the protein backbone RMSD and the ligand's Distance to Bound Configuraiton (DBC) during a 20ns simulation of the solvated lysosyme-phenol. The DMTS-NC integrator uses $1.25-5$fs time steps.