Secondary-data-driven and interpretable machine learning for experimental band-gap prediction in inorganic semiconductors

A secondary experimental dataset containing 6,354 band-gap records was curated by removing invalid observations, consolidating repeated measurements, normalising chemically equivalent formulas, and applying an experimental-consistency criterion. This resulted in 2,264 unique compositions. Composition-based descriptors were constructed from elemental and periodic-table properties. Linear Regression, Random Forest, and XGBoost models were evaluated using development-set cross-validation and a locked held-out test set. Chemical-system-disjoint validation, SHAP interpretation, bootstrap stability analysis, and curation-threshold sensitivity analyses were also performed. The final XGBoost model achieved a cross-validation MAE of 0.459 eV and a locked-test MAE of 0.434 eV, RMSE of 0.670 eV, and R2 of 0.770. Chemical-system-disjoint cross-validation produced an MAE of 0.485 eV. In the locked test set, materials from unseen chemical systems had a higher MAE of 0.502 eV than materials from seen systems. SHAP analysis identified electronegativity-related descriptors as the strongest predictors, and feature rankings were highly stable across bootstrap resampling. The study provides a reproducible and interpretable framework for composition-based prediction of experimental band gaps in inorganic semiconductors. The model showed reasonable generalisation across chemically distinct systems, although prediction remained more difficult for high-band-gap and chemically novel materials.