Learning Solvation: The Transition from Machine Learned Potentials of Bulk Solvent to Aqueous Solution

Due to its ubiquity and importance in daily life, highly accurate simulations of liquid water are critical to understand a variety of phenomena. With the onset and fast spread of machine learning methods across disciplines, development of neural network force fields that can capture the desired accuracy with reasonable computational efficiency is possible. Critically, neural networks can learn dynamics of ab initio simulations using high accuracy exchange-correlation function, allowing ab initio accuracy of simulations on a timescale not feasible for true ab initio methods. Recent work has demonstrated the efficacy of such methods when learning bulk structure and dynamics during simulation, but work remains to extrapolate these methods towards accurately simulating salts in solution in the dilute limit.

In particular, structural affects the solute molecule has on the surrounding solvent change the water-water interactions to a degree that is measurable in the dynamic properties of the solvent. Accounting for these deviations is important when constructing a force field that accurately represents both solvent-solvent interactions as well as the solute-solvent and solute-solute interactions. Here, we will present results studying the transition from learning a force field for bulk water to the learning of a force field for solvated sodium chloride in water using deep neural networks as implemented in DeePMD.