Explainable Artificial Intelligence for Multi-Performance Prediction and Sustainable Optimisation of Reinforced Concrete Using Recycled and Supplementary Cementitious Materials

Abstract

The growing demand for concrete infrastructure has intensified concerns regarding the consumption of natural aggregates, generation of construction and demolition waste, and environmental impacts associated with Portland cement production. Recycled aggregate concrete (RAC) incorporating supplementary cementitious materials (SCMs) provides an attractive pathway toward resource-efficient and lower-carbon concrete; however, the simultaneous incorporation of recycled aggregates and multiple SCMs creates complex interactions among mixture constituents that are difficult to capture using conventional empirical approaches. This study develops an explainable artificial intelligence framework for the multi-performance prediction and sustainable optimization of concrete mixtures containing recycled coarse aggregate and supplementary cementitious materials. A computational dataset comprising 28 concrete mixture designs and three replicate observations per mixture was established, resulting in 84 observations. The investigated mixture variables included recycled coarse aggregate replacement, total SCM replacement, silica fume content, cement content, water-to-binder ratio, aggregate proportions, and superplasticizer dosage. Five performance indicators were considered: 28-day compressive strength, split tensile strength, flexural strength, slump, and water absorption. Nine machine-learning algorithms were benchmarked using grouped five-fold cross-validation to avoid information leakage among replicate observations. The evaluated models included linear regression, Ridge, Lasso, Elastic Net, support vector regression, k-nearest neighbors, random forest, gradient boosting, and extra trees. Ridge regression provided the highest coefficient of determination for compressive strength (R² = 0.971), flexural strength (R² = 0.968), slump (R² = 0.852), and water absorption (R² = 0.962), whereas Lasso provided the highest performance for split tensile strength (R² = 0.964). The corresponding mean absolute errors were 1.317 MPa, 0.196 MPa, 5.260 mm, 0.150 MPa, and 0.150 percentage points, respectively, while the MAPE values remained between 2.27% and 6.13%. The results demonstrate that relatively interpretable regularized regression models can provide strong predictive performance for the present structured mixture dataset. The proposed framework establishes a transparent pathway for linking mixture composition with mechanical, fresh-state, and durability-related performance and provides a foundation for sustainable mixture optimization using recycled aggregates and SCMs.