Thermodynamics of Water and Ice from a Fast and Scalable First-Principles Neuroevolution Potential

Machine learning potentials enable molecular dynamics simulations to exceed the size and time scales that can be accessed by first-principles methods like density functional theory, while still maintaining the accuracy of the underlying training data set. However, accurate machine learning potentials come with relatively high computational costs that limit their ability to predict properties requiring extensive sampling, large simulation cells, or long runs to converge. Here, we have developed and tested a neuroevolution-potential model for water trained to hybrid-dispersion-corrected density functional calculations. This model exhibits accuracy and transferability comparable to state-of-the-art machine learning potentials but at a much lower computational cost. As a result, it enabled us to compute well-converged thermodynamic averages and fluctuations. This allowed us to assess the ability of our model to reproduce several thermodynamic properties of water and ice as well as the anomalous behavior of water density, heat capacity, and compressibility. The efficient GPU acceleration of our model and its capability to reproduce water thermodynamics in good agreement with experiments make it suitable for simulating phase transitions and slow dynamic processes in aqueous systems.