Evaluating robustness of machine learned force fields with enhanced sampling methods

Molecular Dynamics simulations are increasing its use and applications in materials science and engineering. For that, Machine Learning force fields have emerged as a useful tool to have ab initio accuracy in energy and forces while keeping the speed of a classical simulation. However, validation for such potentials are not sufficient to demonstrate the effectiveness of the simulation as the evaluation only involves the force and energy accuracies while overlooking the stability of systems. Here we explore the robustness of different machine learned force fields for different systems and machine learning models such as DeePMD, Graph Neural Network force fields, Gaussian Approximation Potential (GAP) by performing molecular dynamics simulations and evaluating free energy landscapes as function of appropriate collective variables. The simulations are performed with PySAGES, a python library for advanced sampling simulations, using either ASE or JAX-MD as backends. We showcase how some models are susceptible to exhibit undesired behavior such as bonds breaking, while others maintain the most relevant set of features that are needed to take advantage of them as widely used force fields. We show how to use PySAGES as a tool for quickly evaluating the soundness of ML force fields.