Application of Machine Learned Potentials to Bulk and Interfacial Aqueous Solutions
Samuel Varner, California Institute of Technology
Understanding the interfacial behavior of aqueous solutions is crucial, as many natural processes depend on the interactions at the boundaries between water and other phases, such as air or oil. A notable example is the catalysis of reactions on atmospheric water droplets, driven by the internal electric field at the water interface. Modeling sharp interfaces between different phases requires rigorous approaches due to differences in physical properties like dielectric constant, density, electron density, polarity, and chemical interactions. Accurate modeling of even simple systems, such as the air-water interface, remains challenging and is an active research area. Historically, computational methods have faced a trade-off between accuracy and speed. While ab-initio molecular dynamics (AIMD) simulations offer high accuracy, they are computationally intensive, limiting them to small systems and short timescales. Conversely, classical molecular dynamics (CMD) simulations are faster but rely on empirical force fields that often accurately describe only a few properties. Machine-learned interatomic potentials (MLIPs) bridge this gap by using neural networks trained on density functional theory (DFT)-level force and energy data, achieving both high accuracy and efficiency. We employ the NequIP code to train MLIPs based on DFT calculations of the air-water interface, enabling us to describe both structure and dynamics at the interface with high accuracy. Additionally, we investigate aqueous solutions containing sodium halide salts, offering new insights into the longstanding question of halide adsorption and depletion at the air-water interface. Finally, we combine our MLIP with state-of-the-art enhanced sampling techniques available in PySAGES to elucidate the mechanistic details and temperature dependency of water autoionization, revealing the collective motion of water molecules in transition states and multiple key metastable states.