Understanding the solvent effect on ion-pair dissociation at the air-water interface by machine learning interatomic potential

Solvent interactions play a key role in ion-pair formation and dissociation at the air–water interface, but capturing these effects requires ab initio accuracy and explicit solvent coordinate, which are often beyond standard DFT simulations. Machine Learning Interatomic Potentials (MLIPs) offer a promising solution, though generating diverse training datasets for MLIP training remains challenging. Here, we show that initial sampling by classical simulations followed by ab initio optimization can efficiently produce MLIPs capable of capturing the influence of collective solvent coordinate. Using Ca2+..SO42– dissociation as a model, we demonstrate that an accurate estimation of dissociation free energy and kinetics requires an explicit solvent coordinate and ab initio level of theory. In contrast to the classical force field, an ab initio interactions stabilize the solvent shared ion-pair at the air-water interface by 3 kcal/mol. Relative to the bulk, interfacial solvation increases dissociation free energy and slows the dissociation rate, while formation energy barriers remain largely unchanged. These results show that MLIPs can reliably capture solvent effects with ab initio accuracy for ion-pair thermodynamics and kinetics.