Using explainable machine learning to predict compressive strength of blended concrete: a data-driven metaheuristic approach

In this study, we use highly developed machine learning techniques to accurately estimate the Compressive Strength (CS) of blended concrete, considering its composition, including cement, SCMs (Ground Granulated Blast Furnace Slag (GGBFS) and Fly Ash (FA)), water, superplasticizer, fine/coarse aggregate, and curing age. In addition to these, we examine an array of models, including XGBoost, Decision Trees (DT), Deep Neural Networks (DNN), and Linear Regression (LR). Among them, XGBoost has the best performance in every category. We use the Bayesian optimization method for hyperparameter fine-tuning to improve forecast accuracy. Our in-depth examination demonstrates the better predictive skills of ensemble models like RF and XGBoost over LR, which is limited in capturing data complexity beyond linear relationships. With an R² of 0.952, RMSE of 4.88, MAE of 3.24, and MAPE of 9.94%, XGBoost performs noticeably better than its rivals. Using SHAP analysis, we determine that curing age, water content and cement concentration are the main factors influencing the model’s predictive capacity, with the contributions of superplasticizer and fly ash being minimal. Curing age and cement content have an interesting positive association with CS, but water content has a negative link with CS. These results highlight the value of machine learning, especially the effectiveness of XGBoost, as a potent device for forecasting the CS of mixed concrete. Additionally, the knowledge gained from our research provides designers and researchers in concrete materials with useful direction, highlighting the most important factors for compressive strength. Future studies should work toward additional optimization by attempting to verify these models across a wider variety of concrete compositions and test settings.