Predicting the flexural behavior of reinforced concrete beams with carbon fiber-reinforced polymer and recycled aggregates using KAN

Abstract

The integration of Recycled Concrete Aggregates (RCA) and Carbon Fiber-Reinforced Polymer (CFRP) in reinforced concrete beams offers a sustainable structural solution. However, the non-linear interactions between heterogeneous materials – including varying recycled aggregate ratios, silica fume, and fly ash– pose significant modeling challenges for traditional computational methods. This study proposes a novel predictive framework based on Kolmogorov-Arnold Networks (KANs) to estimate the full load-displacement response of these composite beams. Unlike conventional Multi-Layer Perceptrons (MLP), KANs employ learnable B-spline activation functions on edges, offering superior approximation of complex constitutive laws. Utilizing an experimental dataset of 4.851 samples, the model architecture – specifically grid size and spline order– was optimized via the Tree-structured Parzen Estimator (TPE) algorithm. The results demonstrate that the optimized KAN model achieves a Coefficient of Determination (R²) of 0,974 on independent test data, reducing the Mean Squared Error (MSE) by 80,9% compared to the baseline configuration. This study confirms that KANs provide a robust, high-precision alternative to standard "black-box" models for structural reliability assessment in sustainable infrastructure.