Machine learning techniques to trace the fire response of concrete filled steel tubes

This paper presents a new approach for predicting and understanding the fire response of concrete filled steel tubes (CFSTs) using explainable and symbolic machine learning. By leveraging unsupervised and supervised learning techniques, we develop a number of ML models to understand the fire response of CFSTs, predict their failure modes, mechanical and thermal responses, derive new design equations for CFST behavior under fire, and optimize the design of such columns for improved fire resistance. Our methodology includes clustering for identifying structural performance patterns, regression and classification models for failure prediction, and symbolic regression for generating interpretable models that offer insights into the underlying mechanics. More specifically, the clustering analysis revealed three distinct structural performance patterns among the CFST columns (namely, those governed by the material strength, the geometric properties of the tube, as well as a combination of the magnitude of the loading conditions and boundary conditions). Further, regression and classification models were developed for failure prediction, achieving an accuracy of 88% in predicting buckling and crushing failure modes. Extensive evaluation against existing standards reveals our approach's advantages in accuracy and predictability, with the CatBoost model predicting rebar and core temperature with an accuracy of 95%. This work presents a significant step toward enhancing fire-resistant design through ML-driven discovery, thereby improving fire safety and performance.