From Random Splits to Unseen Solvents: Quantum-Chemistry-Derived Surface Descriptors Improve Solubility Prediction

Solvent screening often requires predictions for solvents that are absent from the training data. Using 103,262 aligned BigSolDB v2.1 records, we constructed a fixed solvent-level OOD benchmark in which five solvent identities were excluded from model training, validation, early stopping, SHAP ranking, and feature selection. A compact RDKit two-dimensional baseline reaches an IID R2 = 0.954 ± 0.001 under random splitting but falls to an OOD R2 = 0.541 ± 0.019 on the held-out solvents. Under the same benchmark, adding Gasteiger charge proxies and monomer-level GFN2-xTB/COSMO σ-moment surface descriptors raises the OOD R2 to 0.756 ± 0.007. A validation-selected 47-feature representation that also includes deterministic solute-solvent compatibility features reaches an OOD R2 = 0.788 ± 0.008, an MAE = 0.413, and an RMSE = 0.574 log10 S unit under the 20-seed protocol. The improvement depends on solvent identity: DMSO, NMP, and cyclohexane show the largest error reductions, ethylene glycol changes little, and n-propyl acetate is slightly worse than the electrostatic baseline. Learner controls and pair-xTB calculations are used as supporting checks; optimized 1:1 dimers and small 1:2/2:1 clusters do not improve the fixed-OOD result. MICA v0.2 distributes the frozen 20-seed ensemble for prepared-descriptor inference with ensemble-disagreement summaries, applicability-domain flags, and solvent-ranking reports.