Nuclear Quantum Effects on the Melting Properties of First-Principles Water Models

The density change between solid and liquid at melting, and the density maximum of the liquid above the melting point are two of the properties that make water the most important liquid for life. It is of interest to assess the accuracy of first-principles water models in predicting these properties when nuclear quantum effects are included. Here, we report path integral molecular dynamics simulations driven by deep neural network (DNN) models trained on data generated with the SCAN, SCAN0, revPBE-D3, and revPBE0-D3 functionals as well as with the MB-pol water potential. We find that the DNN model based on MB-pol agrees well with experiment on the isotope effect on the melting temperature, the density discontinuity, and the density maximum, while the models based on DFT approximations display larger deviations from experiment and may fail to predict correctly the qualitative trends. We attribute the discrepancies with experiment of the DFT-based models to overestimation of the hydrogen bond strength.