Improved prediction of geopolymer concrete compressive strength using metaheuristic-enhanced extreme learning models with combined SHAP and LIME for sustainable construction

The building sector is a major contributor to global carbon dioxide emissions, underscoring the need for environmentally sustainable construction materials. Geopolymer concrete offers an eco-friendly alternative to conventional cement-based concrete. Because experimental determination of its mechanical properties is resource‑intensive, this study develops seven innovative hybrid extreme learning machine (ELM) models, each enhanced with a different metaheuristic optimization algorithm, to improve prediction of compressive strength in geopolymer concrete. These models are benchmarked against a standard ELM and multiple linear regression (MLR) baseline. To support interpretability, SHapley additive exPlanations (SHAP), Local Interpretable Model Agnostic Explanation (LIME), and SelectKBest are applied to rank feature importance, demonstrate variable interactions, and quantify relationships. Model performance is evaluated using mean absolute error (MAE), root mean squared error (RMSE), correlation coefficient (R), Nash-Sutcliffe efficiency coefficient (NSE), Kling-Gupta efficiency (KGE), and Willmott’s Index (WI). The results reveal that the ELM combined with invasive weed optimization (ELM-IWO) demonstrates the highest accuracy among all models with RMSE = 2.928 MPa, R = 0.964, and MAE = 2.261 MPa. Strong performance is also obtained by ELM‑GWO (RMSE = 3.031 MPa, R = 0.963, MAE = 2.260 MPa) and ELM‑MFO (RMSE = 3.049 MPa, R = 0.963, MAE = 2.433 MPa), while the MLR baseline shows the weakest performance (RMSE = 3.974 MPa, R = 0.926, MAE = 3.105 MPa). The findings confirm the effectiveness of metaheuristic-enhanced ELM models for accurate compressive strength forecasting when appropriate variables are incorporated. The outcomes offer practical value for reducing laboratory workload and supporting the development of eco‑friendly construction materials through more efficient prediction processes.