Beyond Energies and Forces: Systematic Evaluation of Machine-Learned Potentials for Solid–Liquid Interfaces

Abstract Atomistic simulations have proven to be a powerful tool in obtaining chemical insights for complex solid–liquid interfaces. However, studying solid–liquid interfaces with first-principles methods is a challenging task due to the underlying computational limitations. Empirical force fields have been extensively applied as an alternative; however, they are limited by the accuracy of the underlying force field. Machine learning potentials (MLPs) have emerged as a powerful solution to overcome this accuracy vs computational cost trade-off. Developments in ML algorithms and molecular representations have led to accurately simulating diverse complex molecular systems, including solid–liquid interfaces. The accuracy of the MLP is currently determined by the errors in energies and forces of the system and by observing structural properties. However, such metrics are insufficient in capturing the true insights into interfaces. In this study, we present a workflow for constructing robust MLPs for solid–liquid interfaces and demonstrate their accuracy beyond energy and force errors with evaluating structural and dynamical properties. Through a systematic dataset construction, we provide valuable insights into developing robust and accurate MLPs, their failure modes, and solutions to overcome their shortcomings. With this work, we aim to provide a guide in building accurate, transferable, and robust MLPs for solid–liquid interface systems.