Machine learning efficiently unearths behavior of water

Researchers have traditionally modeled the phase transition of water using classical, semi-empirical models. However, many classical water models have limited flexibility, polarizability, and transferability.

Sanchez-Burgos et al. combined machine learning with an accurate quantum chemical method to comprehensively simulate the behavior of water at its liquid-vapor phase transition.

“Machine learning models provide an efficient venue for prediction of properties that require large system sizes and long timescales,” co-author Maria Muniz said.

The authors’ method was based on SCAN, a quantum chemistry method with demonstrated ability to reproduce the phase diagram of water.

The team performed direct coexistence simulations, in which the liquid and the vapor phases coexist at an interface, at constant temperature and volume. The team then analyzed the density profile to calculate the densities of both the liquid and vapor phases.

For the bubble nucleation study, the team used the Seeding approach, in which a bubble is already inserted into the liquid water phase. Then the system is spontaneously equilibrated in the canonical ensemble, and the critical bubble that is formed is tracked over time.

While the model required a shift of 40 degrees Kelvin to match experimental values, it predicted lower surface tension and a much higher bubble nucleation than that predicted by a well-known classical model. Furthermore, similar to the classical model, the authors found a preferential orientation of interfacial water molecules, with the hydrogen atom exposed to the vapor phase.

Co-author Ignacio Sanchez-Burgos said further work could include liquid and vapor data, since, despite its success, the current model was only trained on ice and liquid phases of water.

Source: “A deep potential model for liquid-vapor equilibrium and cavitation rates of water,” by Ignacio Sanchez-Burgos, Maria Carolina Muniz, Jorge R. Espinosa, and Athanassios Z. Panagiotopoulos, Journal of Chemical Physics (2023). The article can be accessed at https://doi.org/10.1063/5.0144500 .

This paper is part of the Machine Learning Hits Molecular Simulations Collection, learn more here .