Interpretable prediction of failure strength in high-strength steel T-joints using symbolic regression

This paper presents a novel machine cognition framework to formulate the axial failure strength of high-strength steel T-joints, addressing the limitations of conventional design models and the scarcity of experimental data. To overcome data scarcity, a hybrid synthetic data generation framework combining bootstrap resampling with feature-specific Gaussian perturbation, bounded by geometric validity conditions of HSS T-joints, and kernel density estimation expands the dataset from 147 to 400 samples. Unlike conventional resampling that relies on exact replication, the proposed approach preserves inter-variable correlations and statistical distributions while ensuring that all generated samples remain within structurally admissible parameter ranges. The framework employs symbolic regression to develop closed-form mathematical expression that achieves exceptional predictive performance, with a coefficient of determination of 0.99. In addition to predictive accuracy, the methodology integrates explainable artificial intelligence to enhance interpretability and explainability, providing physical insights into the influence of geometric, material, and welding parameters. These insights highlight width ratio, brace wall thickness, and weld size as dominant predictors of joint capacity. In contrast to the resource- and time-intensive laboratory tests, the derived expression provides fast, computationally lightweight, and practically scalable tools for capacity estimation in early-stage design. The proposed hybrid framework, therefore, represents a transparent, trustworthy, and efficient artificial intelligence solution for structural analysis and design.