A unified structure-based framework for honest, IL-disjoint prediction of ionic-liquid properties

<jats:p>Ionic liquids (ILs) are promising electrolyte platforms, but their combinatorial chemical space makes exhaustive measurement impossible. Existing structure–property models for ILs rely on expensive descriptors and point-wise validation that overstates extrapolation accuracy. I assembled 11,511 experimental records for 1,658 ion pairs from five public sources and built an open, structure-based framework predicting conductivity, density, viscosity, and melting point from SMILES and temperature alone. Under strict IL-disjoint cross-validation, gradient boosting reaches R² = 0.552 for conductivity and R² = 0.828 for density, versus R² = 0.908 under point-wise splitting—quantifying the inflation inherent in conventional validation. SHAP identifies temperature, anion lipophilicity, and cation size as dominant drivers; data density, not model capacity, is the binding constraint for extrapolation.</jats:p>