Mechanical properties of heavyweight concrete made from metakaolin and granite industry waste using ANN-based model

• The study successfully incorporates metakaolin (MK) as a partial cement replacement and granite waste (GW) as a full river sand substitute in heavyweight concrete, promoting the use of industrial by-products for eco-friendly construction. • A 20% replacement of cement with MK provided the highest compressive strength, attributed to improved pozzolanic reactions, despite a reduction in workability due to increased internal friction from MK and GW. • Four ANN-based models were developed, with the firefly algorithm-optimized ANN (ANN-FF) achieving the best performance (R² = 0.9631 training, 0.9329 testing), enabling accurate prediction of compressive strength. • A graphical user interface (GUI) was created for real-time strength prediction, and SHAP analysis highlighted water-to-binder ratio, curing age, and MK content as the most critical factors—enhancing interpretability and facilitating data-driven mix design. The growing demand for sustainable construction has increased interest in utilizing industrial by-products and artificial intelligence (AI) to improve concrete performance. This study investigates the mechanical and workability properties of heavyweight concrete incorporating metakaolin (MK) and granite industry waste (GW) as replacements for cement and natural sand, respectively. Sixteen concrete mixes comprising both normal weight (NW) and heavyweight (HW) types were tested, with MK replacing cement at 0%, 10%, 20%, and 30% by weight, and GW fully substituting river sand. Results showed that higher MK and GW contents reduced slump due to increased internal friction, while a 20% MK replacement yielded the highest compressive strength, attributed to enhanced pozzolanic activity. To predict compressive strength, four artificial neural network (ANN) models were developed, including three hybrid versions optimized using metaheuristic algorithms: firefly algorithm (FF), genetic algorithm (GA), and differential evolution (DE). Among them, the ANN-FF model delivered the best performance, with an R² of 0.9631 for training and 0.9329 for testing, along with the lowest error metrics. SHapley Additive exPlanations (SHAP) analysis identified the water-to-binder ratio, curing age, and MK content as the most influential variables. To support practical application, a graphical user interface (GUI) was developed to enable real-time strength prediction based on mix design parameters. The integrated framework that combines experimental testing, machine learning, and explainable AI improves predictive accuracy, reduces the need for extensive laboratory trials, and promotes data-driven mix optimization for environmentally responsible construction.