An Interpretable Solute–Solvent Interactive Attention Module Intensified Graph-Learning Architecture toward Enhancing the Prediction Accuracy of an Infinite Dilution Activity Coefficient

Abstract

The infinite dilution activity coefficient (γ ∞ ) is a significant thermodynamic property for phase equilibrium prediction. Herein, a solute–solvent interactive attention module is proposed to intensify the graph-learning architecture for construction of an accurate predictive model for γ ∞ . The interactive attention module can adaptively capture the intermolecular interactive information between solute and solvent. The final features obtained by the graph-learning architecture include overall information on the intra- and inter-molecular features and temperature-dependent parameters, which are fed into the dropout deep neural network to make predictions. Multiview analysis of the model performance demonstrates that the proposed predictive architecture exhibits superior accuracy and reliability compared to the competitive model. Furthermore, the results prove that the valuable chemical knowledge learned through the proposed attention module contributes to improving the precision and interpretability of the model. As such, the proposed ln γ ∞ predictive architecture could provide a reliable tool for green solvent screening and actual separation process development.