Assessment of ultimate load and failure modes in CFRP-steel joints using Kolmogorov-Arnold Networks

The effectiveness of externally bonded carbon fibre-reinforced polymer (CFRP) for strengthening steel structures hinges on numerous factors. Accurate prediction of ultimate load capacity, failure modes, and the contributions of key influencing features in CFRP-steel joints is essential for ensuring structural integrity. This study applies machine learning techniques, comparing conventional Artificial Neural Networks (ANN) with the novel Kolmogorov-Arnold Networks (KAN), on a dataset comprising 313 experimental observations to build regression and classification models. Results show that KAN surpasses ANN in ultimate load prediction, attaining an R 2 (coefficient of determination) of 0.97 on the training set and 0.95 on the test set, whereas ANN displays inferior accuracy and a propensity for overfitting. In failure mode classification, KAN achieves higher precision across all categories. Feature importance was evaluated using the post-hoc explainable artificial intelligence method SHapley Additive exPlanations (SHAP) for ANN and KAN's inherent interpretability. While ANN with SHAP provides dataset-specific rankings that may be susceptible to noise, KAN uncovers more generalisable functional relationships. For engineering optimisation, integrating both methods are recommended to harmonise data-driven insights with robust generalisation, thereby reducing biases from over-reliance on a single approach. KAN's strengths in handling data scarcity and improving interpretability offer critical advantages for engineering joint analyses, delivering reliable guidance for safety-critical structural predictions.