Machine‐Learning Surrogate Modeling of DFT ‐Calculated Binding Energies for Ionic Liquid Ion‐Pairs

Abstract

Ion–pair binding energy is a fundamental property that governs the physicochemical behavior of ionic liquids (ILs), influencing ion dissociation, transport, interfacial behavior, and field‐driven extraction processes across a wide range of applications. Accurate prediction of this quantity is therefore essential for rational IL design. While density functional theory (DFT) delivers precise energetic characterization of ion pairs, its computational demands render it impractical for high‐throughput evaluation of ionic liquid (IL) candidates. Here, we present a machine learning (ML) approach for direct prediction of DFT‐derived ion‐pair binding energies from SMILES strings of component ions. To resolve the mapping from molecular structure to binding energy, we considered seven machine learning surrogates: (1) kernel ridge regression (KRR), (2) Gaussian process regression (GPR), (3) random forest (RF), (4) XGBoost, (5) LightGBM, (6) CatBoost, and (7) graph neural network (GNN), trained on a set of 100 ion pairs and tested on a held‐out set of 20 ion pairs. The gradient‐boosted trees (XGBoost, LightGBM, and CatBoost) achieved the best predictive performance, yielding an average coefficient of determination () of about 0.844 and an average Spearman's rank correlation coefficient () of about 0.879. The results from this study highlight the strong potential of descriptor‐driven ML approaches to estimate ion‐pair binding energies, establishing their utility as computationally efficient alternatives to DFT for preliminary screening of ionic liquids across varied application domains.