Physics-Informed Site-Specific Molecular Interaction Learning Framework for Liquid–Liquid Phase Diagram Reconstruction and Extraction Process Design

Abstract

Liquid–liquid phase behavior is governed by the collective balance of intermolecular interactions and is central to predicting component partitioning and designing extraction-based separation processes. Existing thermodynamic and data-driven approaches can predict phase-equilibrium properties, but provide limited resolution of the molecular-interaction-to-phase-topology relationship underlying coupled binodal boundaries and tie-line families. Here, PSMI is proposed as a physics-informed site-specific molecular interaction learning framework that reconstructs ternary liquid–liquid phase diagrams directly from molecular structures. PSMI learns chemically meaningful intermolecular interaction sites through molecular representation learning and integrates chemical-potential consistency regularization to guide thermodynamically reliable phase reconstruction. The framework predicts coexisting extract and raffinate compositions along continuous phase paths and enables reconstruction of binodal boundaries and tie-line families. Across benchmark tests, PSMI accurately predicts coexisting extract and raffinate compositions and recovers phase-diagram topology, including both binodal shapes and tie-line directions. Interpretability analysis shows that PSMI identifies transferable molecular drivers underlying liquid–liquid phase separation, including chemically meaningful interaction patterns. Industrial case studies, together with a locally deployable web application, further demonstrate the practical utility and accessibility of PSMI for solvent screening and extraction process design. By resolving site-specific intermolecular interactions together with their macroscopic phase-partitioning consequences, PSMI provides a molecular-level perspective on how local chemical environments contribute to liquid–liquid phase behavior. This connection between localized molecular interactions and continuous phase-diagram topology provides a framework for understanding and designing complex liquid–liquid separation systems.