Bridging the Gap Between Machine-Learning-Powered Simulations and Experiments for the Free Energies of Liquids and Solids

We have recently developed an approach for computing chemical potentials and free energies of both liquid and solid systems at the density functional theory (DFT) level of accuracy using machine learning interatomic potentials (MLIPs). Unlike conventional molecular dynamics (MD) simulations, from which we are unable to rigorously compute the free energy quantities, our method relies on a series of MD simulations with MLIPs using a rigorous statistical mechanics framework based on thermodynamic integration. This approach has been successfully employed to determine the melting points of molten salts, and we are now expanding its applications for predicting solubility, redox potentials, activity coefficients, and phase diagrams, all at DFT or higher levels of accuracy. These free energy properties have traditionally been overlooked in computational studies due to the challenges of computing them from MD simulations, even though they are often readily available from experimental studies. We will discuss several challenges associated with using MLIPs and how we address them to achieve a robust and efficient workflow. Our work serves to bridge the gap between advanced simulations using MLIPs and experimental efforts to determine the key thermodynamic properties of liquids and solids.