Interpretable Machine Learning for Predicting the Compressive Strength of Waste Glass Powder Concrete

Compressive-strength prediction supports mixture evaluation for waste-glass-powder (WGP) concrete, but models trained on literature data may not transfer well across studies. This study re-analysed a published database of 337 records from 18 sources and compared six machine-learning models using a random 80:20 split and nested leave-one-source-out (LOSO) validation. Extra Trees achieved the best random-test performance (R² = 0.921, RMSE = 3.506 MPa, MAE = 2.376 MPa), but its training R² of 1.000 showed a substantial train–test gap. Under LOSO validation, AdaBoost performed best (R² = 0.444, RMSE = 8.799 MPa), the differences among the top five models were within bootstrap uncertainty, and all models transferred less well to unseen sources. SHAP and accumulated local effects analyses ranked curing age, cement content and water-to-cement ratio as the leading predictors and revealed nonlinear effects of replacement level and aggregate proportions. Glass-oxide contents varied only between studies, so their influence could not be resolved. A graphical interface reports predictions, 95% cross-conformal prediction intervals and sparse-domain warnings; the interval covered 97% of random test records but only 71% of LOSO residuals. The results support mixture screening within the sampled domain, while external validation remains necessary.