Machine Learning Accurately Predicts Water's Melting Properties

This paper investigates the use of machine learning potentials (specifically Deep Potential models) to simulate the melting properties of water and ice, including the melting temperature, density discontinuity, and temperature of maximum density. The study compares different potential models, including those trained on Density Functional Theory (DFT) data and the MB-pol potential, against experimental results. The key finding is that the MB-pol based model accurately reproduces experimental observations, while DFT-based models show discrepancies attributed to overestimation of hydrogen bond strength. This work highlights the potential of machine learning for accurate simulations of complex aqueous systems and provides insights into the limitations of certain DFT approximations.