Real-world mega-project data and particle-packing features drive AI design of marine high-performance concrete

The long-term durability and sustainability of marine high-performance concrete (HPC) are decisive for structures exposed to chloride-rich environments, such as sea-crossing tunnels and bridges. However, most AI-based mix-design studies rely on limited lab-scale mixes and omit aggregate gradation, limiting their applicability in real-world projects. To address this gap, this study develops an AI-driven framework that combines machine learning–based performance prediction with multi-objective optimization, supported by a large-scale field dataset comprising 939 HPC mix designs collected from major marine infrastructure projects, including the Shenzhen–Zhongshan Link and the Huangmaohai Sea-Crossing Passage. The dataset uniquely incorporates full particle-size distribution descriptors—eight cumulative sieve residue percentages and maximum aggregate size—to capture granular packing effects. A random forest (RF) model trained on this dataset accurately predicts key properties including slump, 28-day compressive strength, and 28-day rapid chloride penetration coefficient (R² = 0.90–0.95). Mix proportions are then optimized using the NSGA-II algorithm to minimize cement content and carbon emissions while satisfying strength, workability, and durability requirements. SHAP analysis reveals that particle-size features rank among the top five most influential variables for durability prediction. Experimental validation of five AI-designed mixes (C40–C80) demonstrates excellent performance, with 28-day strengths ranging from 44.4 to 80.5 MPa and chloride diffusion coefficients as low as 1.4 × 10⁻¹ ² m²/s. The results highlight the practical value of integrating field-scale data and explicit aggregate gradation features into interpretable, low-carbon mix design strategies for marine concrete applications. • A dataset of 939 marine HPC mixes was collected from mega-projects. • Full PSD descriptors were used to capture packing effects. • Slump, strength, and durability were accurately predicted by RF models. • Mixes were optimized by NSGA-II for reduced cement and emissions. • Strong performance of AI-designed mixes was experimentally validated.