Physics-informed hybrid machine learning for predicting impact resistance of carbonate natural building stones
Construction and Building Materials, cilt.543, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 543
- Basım Tarihi: 2026
- Doi Numarası: 10.1016/j.conbuildmat.2026.148256
- Dergi Adı: Construction and Building Materials
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Compendex, INSPEC
- Anahtar Kelimeler: Carbonate rock, Hybrid model, Impact resistance, Machine learning, Natural stone, SHAP
- Eskişehir Osmangazi Üniversitesi Adresli: Evet
Özet
The impact resistance (BS EN 14158) of carbonate natural stones has not been addressed by any published machine learning study, despite its direct effect on flooring safety. The target IR = Wt/V algebraically embeds all impact-test variables: a leakage pathway overlooked in prior ML work. We removed eight variables and retained four leakage-free predictors: the coefficient of restitution (RC), Knoop hardness (KH), a multiplicative Static Mechanical Index (SMI_mul), and plate thickness, retained by a three-phase Boruta–backward–forward selection pipeline. The resulting framework, the Physics-informed Hybrid with Compensatory Prediction Principle (PHy-CPP), has three layers. An Extra Trees regressor captures general mechanical rules. An IS-activated sigmoid blend shifts weight to a physics-grounded linear model under extrapolation. Dual Gauss-weighted chemical channels (Fe₂O₃ + RC; LoI) correct residuals of chemically anomalous stones. Performance improved monotonically from the Extra Trees baseline (R² = 0.9045, MAPE = 10.86%) to the full PHy-CPP (0.9623, 7.83%). The dataset comprised 51 observations from 17 stone types (7 marbles, 5 limestones, 5 travertines). Validation used nested 17-fold Leave-One-Group-Out cross-validation with a bootstrap 95% CI for R² of [0.948, 0.972]. Given the small sample (17 groups), all figures use this nested, leakage-free protocol, and the physics-informed layers' incremental gains are treated as directional rather than statistically established. Mapping these predictions onto the published breaking-potential thresholds classified 14 of 17 stone types correctly, with 2 of 3 misclassifications at category boundaries. Code and dataset are openly available on GitHub and archived on Zenodo (DOI: 10.5281/zenodo.19788669).