Improving the Stability of GNN Force Field Models by Reducing Feature Correlation

• DOI: 10.48550/arxiv.2502.12548
• Publication Date: 2025-02-18
• Authors: Yujie Zeng; Wenlong He; Ihor Vasyltsov; Jiaxin Wei; Ying Zhang; Lin Chen; Yuehua Dai
• Keywords: Feature (linguistics); Force Field

Recently, Graph Neural Network based Force Field (GNNFF) models are widely used in Molecular Dynamics (MD) simulation, which is one of the most cost-effective means semiconductor material research. However, even such provide high accuracy energy and force Mean Absolute Error (MAE) over trained (in-distribution) datasets, they often become unstable during long-time MD simulation when for out-of-distribution datasets. In this paper, we propose a feature correlation method GNNFF to enhance stability simulation. We reveal negative relationship between models, design loss function with dynamic coefficient scheduler reduce edge that can be applied general training. also an empirical metric evaluate Experiments show our significantly improve especially data less than 3% computational overhead. For example, ensure stable time from 0.03ps 10ps Allegro model.