Incorporating Long-Range Interactions in Machine Learning Models of Water and Aqueous Electrolyte Solutions

In recent years, atomistic machine learning models have become increasingly popular in molecular simulations, given their ability to combine the accuracy of quantum mechanical representations with the speed and efficiency of classical potentials. These models are capable of learning highly complex and multi-dimensional interactions within a local environment but face challenges in capturing long-range behavior. In this work, we train deep neural networks to interatomic Potential Energy Surfaces to construct many-body classical potentials that accurately represent the structure and dynamics of water and electrolyte systems. We introduce a formalism to represent long-range electrostatics by decomposition of the Coulombic Ewald energy from the local environment representation. We demonstrate the effectiveness of this approach by predicting electrolyte solution properties and vapor-liquid coexistence behavior, which highlights the versatility of our models in representing heterogeneous and charged systems.