A new approach for the prediction of partition functions using machine learning techniques.

Using machine learning (ML), we predict the partition functions and, thus, all thermodynamic properties of atomic and molecular fluids over a wide range of temperatures and pressures. Our approach is based on training neural networks using, as a reference, the results of a few flat-histogram simulations. The neural network weights so obtained are then used to predict fluid properties that are shown to be in excellent agreement with the experiment and with simulation results previously obtained on argon, carbon dioxide, and water. In particular, the ML predictions for the Gibbs free energy, Helmholtz free energy, and entropy are shown to be highly accurate over a wide range of conditions and states for bulk phases as well as for the conditions of phase coexistence. Our ML approach thus provides access instantly to G, A, and S, thereby eliminating the need to carry out any additional simulations to explore the dependence of the fluid properties on the conditions of temperature and pressure. This is of particular interest, for e.g., the screening of new materials, as well as in the parameterization of force fields, for which this ML approach provides a rapid way to assess the impact of new sets of parameters on the system properties.