Graph-neural-network potential energy surface to speed up Monte Carlo simulations of water cluster anions

Regression of potential energy functions stands as one of the most prevalent applications of machine learning in the realm of materials simulation, offering the prospect of accelerating simulations by several orders of magnitude. Recently, graph-based architectures have emerged as particularly adept for modeling molecular systems. However, the development of robust and transferable potentials, leading to stable simulations for different sizes and physical conditions, remains an ongoing area of investigation. In this study, we compare the performance of several graph neural networks for predicting the energy of water cluster anions, a system of fundamental interest in Chemistry and Biology. Following the identification of the graph attention network as the optimal aggregation procedure for this task, we obtained an efficient and accurate energy model. This model is then employed to conduct Monte Carlo simulations of clusters across different sizes, demonstrating stable behavior. Notably, the predicted surface-to-interior state transition point and the bulk energy of the system are consistent with findings from other investigations, at a computational cost three-orders of magnitude lower.