Accelerating phase diagram construction through activity coefficient prediction

Obtaining phase diagrams from molecular simulations remains computationally demanding due to the need for extensive sampling of coexistence conditions and large system sizes. This study demonstrates a novel methodology for efficiently predicting phase behavior in Lennard-Jones mixtures by leveraging machine learning. Here, we train a Gaussian process (GP) model on Kirkwood-Buff Integrals (KBIs) to establish a predictive link between KBIs and activity coefficients-a key thermodynamic quantity encoding deviations from ideality. Through the incorporation of KBI trends, the GP model leads to the prediction of the activity coefficients of two new systems without prior knowledge of their phase behavior, eliminating the need for direct coexistence simulations and significantly reducing computational cost. This framework has broad applicability in computational thermodynamics, offering a scalable strategy for studying complex mixtures with tunable interactions.