Unified axial capacity prediction of FRP-confined concrete across natural and recycled aggregate systems using a Kolmogorov-Arnold Networks based cross-domain transfer-learning framework

Reliable axial capacity prediction of FRP-confined recycled concrete remains difficult because most existing empirical models and code-based formulas were developed for natural aggregate concrete and do not maintain stable accuracy across different recycled aggregate systems. This study aims to establish a unified prediction framework for FRP-confined concrete covering natural aggregate concrete, recycled aggregate concrete, recycled lightweight aggregate concrete, and recycled brick aggregate concrete. A database comprising 1470 circular FRP-confined concrete short-column specimens was compiled from published tests. Nine existing empirical and code-based methods were first assessed, and the results showed clear cross-optimal behavior, indicating that no single conventional model provides consistently reliable prediction across aggregate types. A data-driven framework based on Kolmogorov-Arnold Networks (KAN) and cross-domain transfer learning was then developed. Six machine learning models were benchmarked on the natural aggregate source domain, and KAN was selected as the optimal base model. Seven hierarchical fine-tuning strategies were subsequently designed to transfer learned knowledge to three recycled aggregate target domains. Compared with direct learning, the proposed transfer framework achieved markedly improved test performance, especially for the small-sample target domains, with the best R2[jls-end-space/]values reaching 0.975 for recycled aggregate concrete, 0.999 for recycled lightweight aggregate concrete, and 0.976 for recycled brick aggregate concrete. Parametric response surfaces, SHAP analysis, and ICE curves showed that the transferred models maintained smooth response trends and physically reasonable confinement effects. A graphical user interface was also developed for rapid engineering prediction. The proposed framework provides a practical and interpretable tool for unified axial capacity assessment of FRP-confined concrete across natural and recycled aggregate systems.