Beyond Force Metrics: Pre-Training MLFFs for Stable MD Simulations

Machine-learning force fields (MLFFs) have emerged as a promising solution for speeding up ab initio molecular dynamics (MD) simulations, where accurate force predictions are critical but often computationally expensive. In this work, we employ GemNet-T, a graph neural network model, as an MLFF and investigate two training strategies: (1) direct training on MD17 (10K samples) without pre-training, and (2) pre-training on the large-scale OC20 dataset followed by fine-tuning on MD17 (10K). While both approaches achieve low force mean absolute errors (MAEs), reaching 5 meV/A per atom, we find that lower force errors do not necessarily guarantee stable MD simulations. Notably, the pre-trained GemNet-T model yields significantly improved simulation stability, sustaining trajectories up to three times longer than the model trained from scratch. By analyzing local properties of the learned force fields, we find that pre-training produces more structured latent representations, smoother force responses to local geometric changes, and more consistent force differences between nearby configurations, all of which contribute to more stable and reliable MD simulations. These findings underscore the value of pre-training on large, diverse datasets to capture complex molecular interactions and highlight that force MAE alone is not always a sufficient metric of MD simulation stability.

💡 Research Summary

This paper investigates how pre‑training a graph‑neural‑network‑based machine‑learning force field (MLFF) influences the stability of ab‑initio molecular dynamics (MD) simulations. The authors use GemNet‑T, a directional message‑passing GNN, and compare two training regimes: (1) training from scratch on 10 K configurations of the MD17 aspirin dataset, and (2) pre‑training on a large, chemically diverse subset of the Open Catalyst 2020 (OC20) dataset (2 M structures) followed by fine‑tuning on the same 10 K aspirin data. Both models achieve comparable force mean absolute errors (MAE) of about 5 meV/Å, indicating that the pre‑training does not significantly improve raw force prediction accuracy.

MD simulations are then run for 300 ps (600 k steps, 0.5 fs timestep) using a Nosé‑Hoover thermostat at 500 K and a Velocity‑Verlet integrator. Stability is quantified by the “instability onset time” (t_inst), defined as the first moment when any bond length deviates more than 0.5 Å from its reference value. The scratch‑trained model typically fails after ~95 ps, with some runs breaking as early as 70 ps, whereas the pre‑trained model remains stable for the full 300 ps in all trials—a three‑fold improvement.

To understand the origin of this difference, the authors analyze latent representations, force smoothness, and structural metrics. Pre‑training yields more structured atom and edge embeddings, reduced variance in embedding distances for neighboring configurations, and smoother force responses to small geometric perturbations (force‑smoothness metric reduced by ~35 %). Consequently, the force field exhibits less abrupt force fluctuations during integration, which translates into better energy conservation and temperature control. Structural validation using the pair‑distance distribution function h(r) shows that the pre‑trained model’s h(r) MAE (≈0.018 Å) is substantially lower than that of the scratch model (≈0.045 Å), indicating fewer unphysical distortions such as bond stretching or angle collapse.

The study concludes that large‑scale, multi‑domain pre‑training equips MLFFs with transferable chemical knowledge that mitigates over‑fitting on small datasets and enhances the smoothness and consistency of predicted forces. Force MAE alone is insufficient to guarantee MD stability; complementary metrics such as bond‑length monitoring and h(r) deviations are essential. The findings suggest that pre‑training on diverse datasets is a practical pathway to robust, long‑timescale ML‑driven MD simulations for complex systems like catalysts, polymers, and battery electrolytes. Future work will explore broader cross‑domain pre‑training and regularization techniques to further improve physical fidelity.