A Progressive Multimodel Ensemble and Active Learning Framework for Leakage-Free Multiproperty Prediction of Ionic Liquids

Abstract The ILThermo database, the most comprehensive public collection of ionic liquid (IL) thermophysical measurements, covers fewer than one percent of all possible cation–anion pairs, leaving the vast majority of the IL design space experimentally uncharacterized. We report a machine learning pipeline that simultaneously predicts eight physicochemical and biological properties of ILs: density, dynamic viscosity, surface tension, refractive index, melting point, thermal decomposition temperature, glass transition temperature, and cytotoxicity, directly from molecular structure, without requiring prior experimental measurements for the target compound. Working from 32,455 experimental observations across 4,796 unique cation–anion pairs drawn from the ILPE.db curated data set, we train and evaluate six model classes under a strict pair-level split designed to prevent leakage between training and test sets; our analysis shows a naive row-level split would contaminate 59.6% of test pairs. The sequence spans Ridge regression, Random Forest, two Multi-Task Learning (MTL) neural networks, and two LightGBM variants using Morgan fingerprints or ChemBERTa embeddings as molecular representations. An R2-weighted ensemble of the four best-performing models achieves a mean test R2 of 0.752 across all eight properties (0.951 for density; 0.933 for refractive index). The ensemble is then used to screen 80,000 virtual IL pairs from the ILPE.db virtual library by combined model-disagreement uncertainty and Tanimoto novelty scoring, producing a prioritised experimental schedule that separates into uncertainty-driven and novelty-driven candidate clusters.