Kolmogorov-Arnold network integrated with marine predators algorithm for explainable multi-target prediction of mechanical and durability performance of PET- and RHA-blended sustainable concrete

Concrete incorporating recycled polyethylene terephthalate (PET) and rice husk ash (RHA) offers significant environmental benefits through waste valorization and reduced cement consumption; however, accurately predicting its mechanical and durability performance remains challenging because of the complex nonlinear interactions among constituent materials. This study proposes an explainable multi-target prediction framework based on a Kolmogorov-Arnold Network (KAN) optimized using the Marine Predators Algorithm (MPA) for simultaneous estimation of six key concrete properties: compressive strength (CS), split tensile strength (STS), flexural strength (FS), water absorption (WA), modulus of elasticity (MoE), and rapid chloride permeability test (RCPT). The experimental database consisted of ten PET-RHA concrete mixtures, and MixUp augmentation was applied exclusively to the training subset to enhance data diversity while preserving physically meaningful mix-design relationships. The KAN architecture employs learnable B-spline activation functions that enable efficient representation of nonlinear material behavior and provide intrinsic interpretability. MPA was used to optimize critical hyperparameters, including spline order, grid size, and regularization coefficient, thereby improving model robustness and reducing manual tuning requirements. Model interpretability was further enhanced through Integrated Gradients, which quantified the contribution of individual mix-design variables to each predicted output. The proposed KAN-MPA framework achieved an overall coefficient of determination (R²) of 0.9995 with consistently low prediction errors across all target properties, demonstrating strong agreement between predicted and experimental values within the investigated design space. Feature attribution analysis revealed that cement, PET, and RHA contents were the dominant factors influencing both mechanical and durability performance. The developed framework provides a transparent and accurate tool for sustainable concrete mix evaluation and optimization. Nevertheless, the findings should be interpreted within the scope of the available experimental dataset, and future studies should validate the framework using larger and more diverse datasets to assess its broader applicability.