Bridging Machine Learning and Redlich–Kister Theory for Solid–Liquid Equilibria Prediction of Binary Eutectic Solvent Systems

Abstract

Eutectic solvents (ESs) have gained significant interest in various chemical processes due to a broad spectrum of attractive properties, whereas their rational design is currently still in its infancy. To bridge this gap, Redlich–Kister (RK) theory and machine learning are linked for the solid–liquid equilibria (SLE) prediction of ES systems, which is thermodynamically the cornerstone for ES design. RK theory with two or three parameters is first evaluated by fitting experimental SLE of an extensive ES database, demonstrating that the two-parameter-based one is sufficiently reliable for eutectic behavior correlation. Three machine learning methods, namely, Random Forest, multiple linear regression, and ElasticNet, are developed for relating the parameters of RK theory to the RDKit descriptors of ES components. The SLE predictions from RK theory parametrized by the developed machine learning models are carefully evaluated and further externally examined on several recently reported ES systems.