Machine learning-guided reparameterization of the TIP4P water model for accurate thermal and electrical property predictions

Accurately modeling the thermal and electrical properties of water remains a fundamental challenge in molecular dynamics (MD) simulations due to competing physical mechanisms that govern transport behavior. In this study, we present a machine learning (ML)-guided reparameterization framework for the widely used TIP4P water model to improve the prediction accuracy of key macroscopic properties of thermal conductivity, dielectric constant, diffusion coefficient, and density. We employ an optimized neural network trained on simulation data and extend it with explainable AI (XAI) techniques, particularly Deep Symbolic Optimization (DSO), to uncover interpretable mathematical relationships between molecular parameters and target properties. While conventional ML models often overfit to numerical targets without physical fidelity, our XAI-augmented approach enabled systematic tuning of Lennard-Jones parameters, partial charges, and charge location. The resulting model, TIP4P/XAIe, demonstrates significantly improved accuracy, predicting dielectric permittivity within 10 % of the experimental value, thermal conductivity within 30 %, and the diffusion coefficient within 5 %, while preserving the correct temperature-dependent trends. This work highlights the limitations of purely data-driven approaches and demonstrates the potential of integrating physics-based reasoning with machine learning for next-generation force field development.