Integrating thermodynamic integrations with machine-learning force fields for Predicting entropy of liquids at extreme conditions

Computing liquid entropy is important for predicting thermodynamic properties of materials at extreme conditions. However, it is usually the most challenging energy term to obtain with stuffiest accuracy. We propose a computationally inexpensive ab initio approach for computing liquid entropies with density functional theory (and possibly beyond) level accuracy. It utilizes a two-step integration approach using an intermediate state that enables efficient and effective training of machine-learning force fields at a target thermodynamic state. The method is validated for three systems exhibiting diverse types of bonding, namely, metallic Sn, ionic LiF, molecular CO₂ and polymeric/covalent CO₂.