The dark side of the forces: assessing non-conservative force models for atomistic machine learning

Filippo Bigi, Marcel Langer, Michele Ceriotti

International Conference on Machine Learning 2025 · Oral

This talk delves into the critical challenges and innovative solutions for integrating machine learning models into classical mechanics simulations, particularly in the realm of atomistic molecular dynamics. Presented by Filippo Bigi, Marcel Langer, and Michele Ceriotti, the research addresses a fundamental dilemma in ML-driven physical simulations: the trade-off between computational speed and physical correctness. While directly predicting forces with neural networks (non-conservative models) offers significant speed advantages, it often leads to simulations that violate fundamental physical laws, rendering their results unreliable.

AI review

Bigi, Langer, and Ceriotti present a technically honest and practically motivated study of non-conservative force models in ML-driven molecular dynamics. The core diagnostic contribution — a rigorous characterization of why direct force prediction breaks classical mechanics — is well-executed, and the proposed remedies (hybrid pre-training plus conservative fine-tuning, and Multiple Time Stepping at inference) are sensible and clearly validated. The work is competent, grounded in correct physics, and useful for the ML-for-science community. It does not, however, establish a new theoretical…