Hybrid Approach for Predicting Melting Points in Nonionic Eutectic Solvents Using Thermodynamics and Machine Learning
Dmitriy M. Makarov, Vasiliy A. Golubev, A. M. Kolker
Abstract
In this work, a hybrid approach combining solution thermodynamics and machine learning (ML) methods is presented as a means of estimating solid–liquid equilibria (SLE) in nonionic eutectic solvents. The models were developed based on a data set comprising 141 binary mixtures and 1668 experimental melting points. The semiempirical Associated Solution and Lattice (ASL) method was employed to characterize the SLE in two versions: with one fitting parameter, representing the interchange energy (ASL(ω)), and with two fitting parameters, representing the interchange energy and the heteroassociation constant (ASL(ω′, K )). This work compares models for predicting mixture melting points using direct ML and a hybrid approach. In the hybrid method, ML first predicts the ASL model’s fitting parameters, which are then used to calculate melting points. The single-parameter ASL approach showed better predictive performance than both the two-parameter ASL and direct ML predictions, achieving the lowest average absolute deviation of 8.7 K.