Optimized Machine Learning Models for Predicting Compressive, Tensile, and Flexural Strengths of Multi-Fiber Recycled Aggregate Concrete

The demand for concrete has led to increased use of raw materials and significant waste generation. Recycled aggregate concrete (RAC) offers a viable approach to sustainable concrete; however, the use of weakly bonded mortar on aggregate leads to low strength and crack formation. Fiber reinforcement, specifically hybrid fiber reinforcement combining steel, glass, basalt, and polypropylene fibers, can increase the tensile and flexural properties of RAC. This study developed machine learning models to enable the prediction of hybrid fiber-reinforced RAC’s compressive, splitting tensile, and flexural strength performance; these new models overcome the limitations of previous research, which relied on only one fiber type and regular methods of optimization. Two models (a deep neural network (DNN) and an XGBoost model) were trained and optimized using bald eagle search (BES), particle swarm optimization (PSO), and the Bayesian optimization (BO) algorithm to improve performance. Among the three optimization analyses, PSO-XGBoost achieved the highest accuracy for compressive strength and splitting tensile strength, while BES-XGBoost achieved the highest accuracy for flexural strength. The most significant influences on the compressive strength were curing age and silica fume, while the main drivers of splitting tensile strength and flexural strength were fiber volume and fiber characteristics. The use of SHAP-based methodology with a user-friendly interface further improved the design of RAC mixtures, reducing waste from raw materials, enhancing the structural performance of RAC, and enabling data-driven decision-making in the manufacturing of eco-friendly concrete products.