Accelerating Charge Estimation in Molecular Dynamics Simulations Using Physics-Informed Neural Networks: Corrosion Applications

Abstract Molecular Dynamics (MD) simulations can be utilized to develop a fundamental understanding of the effect that corrosion has on metallic materials in the presence of a salt brine. The use of reactive force fields in classical MD offers the capability to accurately model the formation and breakage of bonds within the aqueous medium and at the interface between the metal and the electrolyte. Additionally, MD protocols can facilitate dynamic partial charge equilibration. Nevertheless, despite their capabilities, MD simulations are unsuitable for modeling the long time scales characteristic of corrosive phenomena given their computationally intensive nature. Hence, there is a critical need for a protocol to decrease the computational expense associated with variable-charge molecular dynamics simulations. This work addresses this concern by developing reduced-order machine learning models to provide accurate and computationally efficient predictions of the charge density in corrosive environments. Specifically, Long Short Term Memory (LSTM) networks are used to assess and forecast the charge density evolution as a function of the atomic environment, represented using Smooth Overlap of Atomic Positions (SOAP) descriptors. Furthermore, this work utilizes a physics-informed loss function to enforce charge neutrality and the electronegativity equivalence. Lastly, even though the protocols are developed in the context of corrosion, they have been formulated in a phenomena agnostic manner, facilitating their utilization for a wide-class of variable-charge interatomic potentials and related applications.