Automatic Refinement of Force Fields Based on Phase Diagrams

Abstract

Exact characterization of phase transitions requires sufficient configurational sampling, necessitating efficient and accurate potential energy surfaces. Molecular force fields with computational efficiency and physical interpretability are desirable but challenging to refine for complex interactions. To address this, we propose a force field refinement strategy with phase diagrams as top-down optimization targets based on automatic differentiation. Using gas-liquid co-existence as a paradigm, we employ an enhanced sampling technique and design a differentiable loss function to evaluate force fields' depiction of phase diagrams. The refined force fields produce gas-liquid phase diagrams matching well with targets for two modeling systems, which confirms our approach as an effective automated force field development framework for phase transition studies.