Hybrid metaheuristic (WOA-HHO) optimized machine learning for compressive strength prediction of nano silica modified self compacting recycled aggregate concrete

This study presents a hybrid experimental machine learning (ML) framework to assess and predict the compressive strength of nano silica (NS) modified self-compacting concrete (SCC) incorporating recycled coarse aggregates (RCA). Experimentally, NS exhibited a clear optimum around NS4 (≈4%). For the reference SCC (RCA0), the 28-day strength increased from 58.6 MPa (NS0) to 63.5 MPa (NS4) (+8.36%), reaching 66.5 MPa at 120 days, while higher dosage (NS6) reduced strength. The same NS4 trend persisted under RCA replacement, improving 28-day strength from 45.2 to 49.3 MPa (RCA25), 31.5 to 35.4 MPa (RCA50), 26.4 to 30.5 MPa (RCA75), and 17.8 to 21.6 MPa (RCA100). Moreover, a comprehensive dataset of 387 samples (current experiments + literature) was developed using eight inputs (cement, water, RCA, Natural Coarse Aggregates (NCA), Natural Fine Aggregates (NFA), superplasticizer, NS, age) to predict compressive strength. Four ensemble models, Random Forest (RF), Light Gradient Boosting (LGB), Extreme Gradient Boosting (XGB) and Categorical Boosting (CB) were independently tuned using Whale Optimization Algorithm (WOA) and Harris Hawks Optimization (HHO) under nested 5x5 cross-validation with a 20% hold-out test. CB delivered the best CV performance (CB-WOA: R 2 = 0.967 ± 0.011 , RMSE = 2.894 ± 0.460 MPa). On the independent test set, CB-HHO achieved the highest accuracy ( R 2 = 0.973 , RMSE = 2.559 MPa, MAE = 1.813 MPa, MAPE = 4.369%). SHAP and PDP-ICE indicated age and RCA as dominant drivers, with diminishing returns for NS beyond moderate dosages and a clearer strength reduction at higher water contents. By enabling higher RCA utilisation while maintaining strength through NS optimisation and data-driven mix design, the proposed framework supports more sustainable SCC production by reducing natural aggregate demand and construction waste, and by limiting trial and error experimentation through accurate predictive modelling.