Systematic comparison between Graph Neural Networks and UNIFAC-IL for solvent pre-selection in liquid-liquid extraction

Solvent selection is a critical decision-making process that balances economic, environmental, and societal factors. The vast chemical space makes evaluating all potential solvents impractical, necessitating pre-selection strategies to identify promising candidates. Predictive thermodynamic models, such as the UNIFAC model, are commonly used for this purpose. Recent advancements in deep learning have led to models like the Gibbs-Helmholtz Graph Neural Network (GH-GNN), which overall offers higher accuracy in predicting infinite dilution activity coefficients over a broader chemical space than UNIFAC. This study presents a systematic comparison of solvent pre-selection using GH-GNN and UNIFAC-IL in the context of liquid-liquid extraction. The original GH-GNN model is extended to simultaneously predict organic and ionic systems. This extended GH-GNN model predicts more than 92 % of the logarithmic IDACs with an absolute error of less than 0.3. By comparison, UNIFAC-based models only achieve such accuracy for less than 65 %. A case study is used involving the ionic liquid ethyl-3-methylimidazolium tetrafluoroborate ([EMIM][BF4]) and caprolactam, relevant for the solvolytic depolymerization of polyamide 6. Results indicate a significant correlation in solvent rankings across both methods, with a Spearman�s coefficient of 0.62, suggesting that deep learning-based models like GH-GNN are viable alternatives for solvent pre-selection. Additionally, chemical similarity metrics, such as Tanimoto similarity, can assess confidence in solvent rankings, allowing users to determine acceptable risk levels in predictions across a vast chemical space.