We have used a deep learning-based active learning strategy to develop ab initio level accurate machine-learned (ML) potential for a solution-phase reactive system. Using this ML potential, we carried out enhanced sampling simulations to sample the reaction process efficiently. Multiple transitions between the reactant and product states allowed us to calculate the converged free energy surface for the reaction. As a prototypical example, we have investigated the Menshutkin reaction, a classic bimolecular nucleophilic substitution reaction (SN2) in aqueous medium. Our analyses revealed that water stabilizes the ionic product state by enhanced solvation, facilitating the reaction and making it more spontaneous. Our approach expands the scope of studying the chemical reaction under realistic conditions, such as explicit solvents at finite temperatures, closely mimicking experiments.
Read full abstract