Explainable artificial intelligence framework for predicting the compressive strength of fly ash-based alkali-activated concrete

Fly ash-based alkali-activated concrete is a promising alternative to ordinary Portland cement concrete because it utilizes industrial by-products and offers potential environmental benefits. However, predicting its compressive strength remains challenging because of the nonlinear relationships between mixture composition, curing conditions, and mechanical performance. Although machine learning methods have been increasingly applied to this task, the integration of comparative ANN modeling, cross-validation, explainable prediction, and practical software implementation within a unified framework warrants further investigation. To address this need, this study develops an explainable artificial intelligence framework for predicting the compressive strength of fly ash-based alkali-activated concrete. A dataset comprising 162 concrete mixtures and 16 input variables was used to develop and compare three models: ANN-LM, ANN-GWO, and ANN-ARO, based on the Levenberg–Marquardt algorithm, Grey Wolf Optimizer, and Artificial Rabbits Optimization, respectively. Model performance was evaluated using the coefficient of determination (R²), mean squared error (MSE), root mean squared error (RMSE), mean absolute error (MAE), mean absolute deviation (MAD), and the a20 index. Ten-fold cross-validation was conducted to examine predictive performance across different data partitions, while SHapley Additive exPlanations (SHAP) was used to interpret the influence of input variables on model predictions. Among the three models, ANN-LM achieved the best testing performance, with R² = 0.9296, RMSE = 0.1879, MSE = 0.0353, and MAE = 0.1390. In ten-fold cross-validation, ANN-LM achieved a mean R² of 0.7419, with mean RMSE, MAE, and MAD values of 0.45492, 0.33599, and 0.32469, respectively, and a mean a20 index of 18.64%. The difference between the testing and cross-validation results indicates that predictive performance varies across data partitions, highlighting the need for further assessment using independent datasets. SHAP analysis identified curing temperature, alkaline activator characteristics, and fly ash characteristics as influential predictors of compressive strength. Finally, a MATLAB-based graphical user interface, GeoMixAI, was developed to facilitate preliminary compressive strength estimation from specified mixture inputs. The main contribution of this study is the integration of comparative ANN evaluation, cross-validation, SHAP-based interpretability, and accessible software implementation within a unified framework, linking predictive performance assessment with interpretation and practical application.