From Local Structure to Thermodynamics and Transport of Water with Machine Learning Force Fields

Abstract

We evaluate machine learning force fields derived from different density functional theory exchange correlation functionals using the full six-dimensional pair correlation function of liquid water, three-body structural descriptors, excess entropy, and transport properties. The predicted microscopic structure and dynamics depend strongly on the underlying functional: neglecting dispersion produces pronounced overstructuring, overly negative excess entropy, and suppressed diffusion. Translational and orientational entropy contributions are tightly coupled and together exhibit a clear relationship with the reduced selfdiffusion coefficient. Among the tested models, RPBE-D3 provides the most consistent agreement with experiment across structural, thermodynamic, and transport properties. The classical SPC/E model serves as an additional reference and displays notable similarities to RPBE-D3, consistent with the comparable Born effective and partial charges of the two models.