Improving prediction accuracy of mechanical strengths in high-performance concrete using advanced metaheuristic-based extreme learning machines: incorporation of SHAP and LIME model explainers

High-performance concrete (HPC) is essential for engineering structures exposed to demanding environments where superior strength, durability, and workability are required. Compressive strength (CS) and tensile strength (TS) are key parameters governing structural design. Accurate forecasting of these properties reduces design costs, shortens development time, and minimizes material waste from extensive laboratory testing. The present research develops an innovative prediction framework utilizing extreme learning machine-Harris hawks optimization (ELM-HHO), ELM-flower pollination algorithm (ELM-FPA), ELM-grasshopper optimization algorithm (ELM-GOA), ELM-biogeography-based optimization (ELM-BBO), and ELM-sine cosine algorithm (ELM-SCA) approaches to forecast CS and TS of HPC. Their performance is compared with standard ELM and multiple linear regression (MLR). SHapley additive exPlanations (SHAP) and Local interpretable model-agnostic explanations (LIME) analyses are employed to examine the governing processes behind CS and TS predictions and to quantify individual input variable influences. Model accuracy is evaluated using several statistical measures, including the correlation coefficient (R), root mean square error (RMSE), Nash–Sutcliffe efficiency (NSE), and mean absolute error (MAE). For CS prediction, ELM-GOA demonstrates superior accuracy with an RMSE of 6.422 MPa and MAE of 4.897 MPa, outperforming ELM-HHO (RMSE: 6.708 MPa, MAE: 5.341 MPa) and ELM-BBO (RMSE: 6.862 MPa, MAE: 5.240 MPa), which rank second and third, respectively. In TS prediction, ELM-GOA again emerge as the leading model, achieving exceptional accuracy with RMSE and MAE values of 0.335 MPa and 0.244 MPa. The second-best performer is ELM-FPA (RMSE: 0.339 MPa, MAE: 0.249 MPa), closely followed by ELM-SCA with marginally higher error values (RMSE: 0.341 MPa, MAE: 0.260 MPa). The research confirms that the implemented parameter sets, when integrated with the developed algorithms, successfully enable accurate forecasting of CS and TS in HPC across various experimental conditions.