FRAMES: A Novel Training Strategy to Improve Molecular Force Fields Using MD Simulation Data
A new training strategy called FRAMES leverages temporal information from Molecular Dynamics (MD) simulations to improve the accuracy of neural network predictors for molecular energies and forces. MD simulations generate time-ordered trajectories of atomic positions that explore the potential energy surface under ensembles like NVE or NVT, unlike geometry relaxations that minimize energy. FRAMES introduces an auxiliary loss function that exploits this temporal data, addressing a gap in current models that typically learn from single atomic configurations. The work is detailed in arXiv preprint 2604.19806, submitted in April 2026.
Key facts
- arXiv:2604.19806v1 introduces FRAMES training strategy.
- FRAMES uses an auxiliary loss function to leverage MD simulation temporal data.
- MD simulations produce time-ordered trajectories under NVE/NVT ensembles.
- Current neural networks often predict from single atomic configurations.
- Accurate energy and force prediction is a core challenge in AI for Science.
- FRAMES aims to improve performance of molecular force field predictors.
- The work explores novel use of MD data when available.
- The paper is categorized as a cross submission on arXiv.
Entities
Institutions
- arXiv