![[2602.14975] Faster Molecular Dynamics with Neural Network Potentials via Distilled Multiple Time-Stepping and Non-Conservative Forces](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.14975] Faster Molecular Dynamics with Neural Network Potentials via Distilled Multiple Time-Stepping and Non-Conservative Forces
Summary
The paper presents a novel approach, DMTS-NC, for accelerating molecular dynamics simulations using neural network potentials, achieving significant efficiency improvements.
Why It Matters
This research addresses the need for faster and more efficient molecular dynamics simulations, which are crucial in fields like materials science and drug discovery. By leveraging neural network potentials, the proposed method enhances simulation stability and reduces computational time, making advanced simulations more accessible.
Key Takeaways
- Introduces DMTS-NC, a method that combines distilled multi-time-stepping with non-conservative forces.
- Achieves 15-30% speed improvements over traditional methods without requiring fine-tuning.
- Ensures robustness in simulations by enforcing physical priors like equivariance under rotation.
Physics > Chemical Physics arXiv:2602.14975 (physics) [Submitted on 16 Feb 2026] Title:Faster Molecular Dynamics with Neural Network Potentials via Distilled Multiple Time-Stepping and Non-Conservative Forces Authors:Nicolaï Gouraud, Côme Cattin, Thomas Plé, Olivier Adjoua, Louis Lagardère, Jean-Philip Piquemal View a PDF of the paper titled Faster Molecular Dynamics with Neural Network Potentials via Distilled Multiple Time-Stepping and Non-Conservative Forces, by Nicola\"i Gouraud and 5 other authors View PDF HTML (experimental) 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, t...
