The compressive strength of pervious concrete is challenging to predict due to complex, non-linear interactions between chemical, mix design, and curing factors. This study presents an integrated data-driven framework combining oxide-level chemical descriptors, advanced ensemble machine learning, metaheuristic optimisation, explainable artificial intelligence, and uncertainty quantification for compressive strength prediction. A dataset of 659 samples from 20 published studies was assembled, incorporating CaO, SiO₂, Al₂O₃, Fe₂O₃, MgO, alkali content, aggregate properties, water content, and curing duration as inputs. Five models were developed and evaluated: Random Forest, AdaBoost, LightGBM, CatBoost, and a hybrid CatBoost-MOWCA. CatBoost achieved the best generalisation performance (R 2 = 0.9272, RMSE = 2.334 MPa), outperforming all models on the independent test set and comparing favourably with published literature reporting R 2 values of 0.85–0.95 for similar systems. SHAP analysis identified curing duration, Al₂O₃ content, and water content as the primary strength drivers. The positive Al₂O₃ effect reflects its pozzolanic role in SCM-blended systems, where it participates in secondary reactions forming calcium aluminate hydrate (CAH) and calcium aluminosilicate hydrate (CASH) phases that densify the microstructure and strengthen the interfacial transition zone . Monte Carlo simulation revealed that epistemic uncertainty dominates at 62.1%, consistent with the heterogeneous multi-study dataset composition. An extended CatBoost model incorporating total cementitious material content as an additional input confirmed complementary predictive value, yielding modest but consistent improvement across all test metrics . A graphical user interface powered by the CatBoost model enables real-time prediction and sensitivity analysis, supporting practical mix design optimisation.
Wani et al. (Sat,) studied this question.