Accurate Simulations of Water and Aqueous Solutions through Fine-Tuned Dispersion-Corrected Density Functional Theory and Machine-Learning Interatomic Potentials

Dispersion-corrected density functional theory (DFT-D) is widely employed to model large molecular systems at an affordable computational cost and to develop machine-learning interatomic potentials (MLIPs), enabling reliable molecular dynamics (MD) simulations of condensed-phase systems. Yet, given a molecular system, the choice of a specific DFT-D model that can achieve the necessary accuracy over an extended range of physicochemical properties and conditions is generally not trivial. Here, we report an effective computational strategy for enhancing the accuracy of standard DFT-D models toward high-level quantum mechanical data and for developing MLIPs preserving the same high fidelity. Taking water as a paradigmatic example, we derive a novel MLIP and demonstrate that its use allows us to accurately predict a wide range of properties in diverse forms, from small clusters to bulk liquid and ice, such as radial distribution functions, fusion/vaporization enthalpies, diffusion constants, and density isobars, capturing remarkably well its peculiar and anomalous behavior, often elusive even to standard first-principle MD simulations. Furthermore, we show how the same computational strategy can be readily extended to treat aqueous solutions. Considering MgCl₂ in water as a test case, we develop a MLIP and use it to predict the metal ion hydration structure and the water exchange dynamics exhibiting a significantly improved agreement with experiments with respect to both standard DFT-D and classical force fields.