Dynamic Training Enhances Machine Learning Potentials for Long-Lasting Molecular Dynamics

Abstract

Molecular Dynamics (MD) simulations are vital for exploring complex systems in computational physics and chemistry. While machine learning methods dramatically reduce computational costs relative to ab initio methods, their accuracy in long-lasting simulations remains limited. Here we propose dynamic training (DT), a method designed to enhance accuracy of a model over extended MD simulations. Applying DT to an equivariant graph neural network (EGNN) on the challenging system of a hydrogen molecule interacting with a palladium cluster anchored to a graphene vacancy demonstrates a superior prediction accuracy compared to conventional approaches. Crucially, the DT architecture-independent design ensures its applicability across diverse machine learning potentials, making it a practical tool for advancing MD simulations.