(451k) Evaluating Robustness of Machine Learned Force Fields with Enhanced Sampling Methods

Molecular Dynamics simulations have become a essential part in materials science and engineering. Machine Learning force fields have emerged as a useful tool to achieve ab initio accuracy in energy and forces while maintaining the speed of classical simulations. However, validating the potentials based only on force and energy accuracies is insufficient as system stability is often overlooked. This study evaluates the robustness of different machine learning models such as DeePMD, Graph Neural Network force fields, and Gaussian Approximation Potential (GAP) by performing molecular dynamics simulations using PySAGES, a python library for advanced sampling simulations, with ASE or JAX-MD backends. The study examines the stability of different systems and evaluates free energy landscapes as a function of appropriate collective variables. The results demonstrate that some models exhibit undesired behavior such as bond breaking, while others maintain the most relevant set of features that are needed to take advantage of them as widely used force fields. PySAGES is shown to be a useful tool for quickly evaluating the soundness of ML force fields.