Optimizing fly ash slag concrete mixtures is a complex material-engineering problem because mixture variables interact non-linearly, making conventional single variable approaches inadequate. This study develops an interpretable prediction-and-optimization framework comprising three contributions: (1) benchmarking five machine-learning algorithms Support Vector Regression (SVR), Artificial Neural Network (ANN), Random Forest, XGBoost, and LightGBM using hyperparameters optimized through Bayesian optimization (Optuna/TPE); (2) applying SHAP analysis to identify the contribution of each mixture variable; and (3) using a Genetic Algorithm (GA) to optimize concrete mixtures for three strength classes based on the best performing surrogate model. Starting from the 1,030-sample UCI concrete strength benchmark dataset, records were screened according to three Indonesian standards (SNI 2847:2019, SNI 2461:2011, and SNI 8297:2016), resulting in 542 samples that satisfied the specified technical limits. Data augmentation was evaluated using six noise levels, with 2% identified as optimal, producing a final dataset of 1,074 samples. Five-fold cross-validation demonstrated that SVR achieved the highest predictive performance (C=193.39, gamma=0.658, epsilon=0.146), with R²=0.939±0.020 and RMSE=3.30±0.45 MPa, followed by XGBoost (R²=0.934), LightGBM (R²=0.931), ANN (R²=0.925), and Random Forest (R²=0.920). SHAP analysis identified cement, age, and superplasticizer as the most influential variables, consistent with cement hydration mechanisms and Abrams’ law. GA optimization produced three mix designs corresponding to target strengths of 15.00, 27.50, and 54.99 MPa for Classes I-III, respectively. The proposed framework provides an accurate, interpretable, and standards-compliant approach for preliminary sustainable concrete mixture design across different structural applications.
Copyrights © 2026