Leakage-Safe Probe-Assisted Contact Angle Prediction Using Nonnegative Surface-Energy Summaries and Physics-Residual Learning

Contact-angle prediction from literature data is vulnerable to target leakage when solid surface-free-energy descriptors are reconstructed using the liquid, which is later treated as the target. We developed a target-masked workflow that removes the target liquid before fitting nonnegative Owens-Wendt-Rabel-Kaelble components by nonnegative least squares and uses that physical prediction to anchor residual learning. A row-level revision audit re-extracted or excluded mismatched legacy sources before all models were retrained. Development used nested source-group cross-validation; the fixed cross-source external confirmation set was excluded from selection. The revised residual model achieved mean absolute errors of 15.5 degrees in nested validation and 13.2 degrees on that confirmation set. Surface-cluster bootstrap supported improvement over physics, whereas source-cluster uncertainty remained substantial. Diagnostics showed source dependence, sparse roughness, and limited strict unseen-liquid support. The method is therefore positioned as an auditable, risk-aware and reproducible materials-screening tool for surfaces with at least two non-target probes, with explicit out-of-distribution risk and refusal conditions rather than universal transfer claims.