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

2025-06-17

AI Summary

This work addresses the instability of machine-learned force fields (MLFFs) in molecular dynamics (MD) simulations—a critical issue where low force prediction error (e.g., mean absolute error, MAE) does not guarantee stable MD trajectories. We propose a pretraining-based solution: leveraging the GemNet-T graph neural network, pretrained on the large-scale OC20 dataset and subsequently fine-tuned on small-sample MD17 tasks. Our study provides the first empirical evidence that force MAE is not inherently correlated with MD trajectory stability. Pretraining significantly enhances physical consistency in force modeling: with only 5 meV/Å force MAE, trajectory stability improves by a factor of three. This work challenges the conventional paradigm that relies solely on force accuracy as the evaluation metric for MLFFs, establishing pretraining as a key pathway to improving their physical robustness and long-term dynamical fidelity.

Technology Category

• Machine Learning: Graph-based Machine Learning
• Natural Language Processing: Safety and Robustness
• Multiagent Systems: Other Foundations of Multi Agent Systems

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Web Mining and Content Analysis: Large pretrained models with web data
• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications

Abstract

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. 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.

Problem

• Improving stability of molecular dynamics simulations with MLFFs
• Evaluating pre-training impact on force field accuracy
• Assessing force MAE limitations in simulation stability

Innovation

• Pre-training GemNet-T on OC20 dataset
• Fine-tuning model on MD17 dataset
• Improving MD simulation stability significantly