A Transferable Four-Tier Decision Architecture for Gas Solubility Prediction in Ionic Liquids

We develop a transferable, physics-inspired feature engineering framework that encodes knowledge of solvation thermodynamics, electrostatic interactions, free-volume effects, and acid–base chemistry directly into the input representation, ensuring that the model learns causal physical relationships rather than data set-specific artifacts. Through a four-step ablation protocol under GroupKFold cross-validation, where generalization is evaluated on entirely unseen ionic liquid (cation + anion)─gas compositional triplets, we identify a Pareto-optimal Thermodynamic + Electrostatic model. Critically, SHAP analysis reveals that the model has autonomously learned a four-tier decision architecture that mirrors established physicochemical hierarchies: a thermodynamic baseline, structure-mediated microtuning, an electrostatic binary switch, and a phase-state classifier. The transferability of this architecture was further evaluated through zero-shot generalization to three entirely unseen gases (C2H4, SO2, and N2), which span distinct regions of the physicochemical interaction spectrum and extend the evaluation beyond the gas chemistries represented during training.