Enhanced mechanical and shielding properties of heavy concrete: A machine learning approach to mix proportion optimization.

Shielding concrete is a multifunctional material that combines structural support with radiation shielding, offering significant practical value. However, inherent functional conflicts between mechanical strength and radiation shielding performance remain challenging to address using conventional mix design approaches. This study prepared 41 concrete mixtures with different aggregates, water-cement ratios, and boron carbide (B4C) contents, and tested their compressive strength (CS) and density. Meanwhile, we performed a quantitative analysis to assess the impact of B4C on the CS and shielding properties of concrete. Monte Carlo simulations with realistic geometries were performed to evaluate the transmission factor (TF) and fast neutron removal cross-section (ΣR), while the linear attenuation coefficient (LAC) and half-value layer (HVL) were calculated using the Phy-X/PSD software based on tabulated photon interaction cross sections. We developed a multi-objective optimization model that integrated Bayesian Optimization (BO), Support Vector Regression (SVR), and the non-dominated sorting genetic algorithm with elite strategy (NSGA-II) to balance CS, ΣR, and density. Model performance was evaluated using nested repeated five-fold cross-validation. The results showed that when the water-cement ratio was 0.35-0.50, CS decreased with increasing B4C content (up to 9%) but remained above the C30 design strength. Moreover, a higher water-cement ratio is associated with a greater critical threshold at which B4C begins to significantly affect CS. The Pareto-optimal solutions indicate that favorable mix designs are concentrated around a B4C content of approximately 4-5% and a water-cement ratio of approximately 0.35-0.40.