Graph Neural Network with Inductive Bias for Robust Molecular Dynamics

Machine learning (ML) force field models for molecular dynamics (MD) simulations often suffer from poor system stability with instabilities such as atom clustering that must be corrected by active learning approaches. However, relationship between the structural and chemical complexity of multi-component systems and the robustness of long-time ML-based MD dynamics has not been studied in detail. We develop a graph neural network (GNN) model for a range of material systems to perform ML-MD simulations with quantum mechanical accuracy but orders-of-magnitude faster. A GNN model is sufficient to ensure robust long-time dynamics in a ‘simple’ system like SiC and Cu. However, we need additional inductive bias, in the form of energy decomposition into 2-body and 3-body terms to generate stable MD trajectories for complex systems such as GeSe₂ and VO₂, which can exist in multiple metastable atomic configurations.

 

Acknowledgement

This work was supported as part of the Computational Materials Sciences Program funded by the U.S. Department of Energy, Office of Science, Basic Energy Sciences, under Award Number DE-SC001460. Simulations were performed at the Center for Advanced Research Computing of the University of Southern California.