Explainable machine learning models for predicting the axial compression capacity of concrete filled steel tubular columns

Abstract

Concrete-filled steel tubular (CFST) columns have been popular in the construction industry due to enhanced mechanical properties such as higher strength and ductility, higher seismic resistance, and aesthetics. Extensive experimental, numerical and analytical studies have been conducted in the past few decades to assess the structural response of CFST columns under various loading conditions. However, there is still uncertainty in predicting the capacity of CFST columns, and most of the current codes are conservative. In this paper, data-driven machine learning (ML) models have been developed to predict the axial compression capacity of rectangular CFST columns. An extensive database of 719 experiments was collected from literature and is randomly used to train, test, and validate the ML models. Seven ML models, namely lasso regression, random forest, Adaptive Boosting (AdaBoost), Gradient Boosting Machine (GBM), Light Gradient Boosting Machine (LightGBM), Extreme Gradient Boosting (XGBoost), and Categorical Gradient Boosting (CatBoost), are evaluated to predict the compression capacity of CFST stub columns under axial load. The performance of the different ML models in predicting the compressive strength of CFST columns is compared by different code equations prevalent in different parts of the world. It is found that LightGBM and CatBoost models performed better with an accuracy of 97.9% and 98.3%, respectively, compared to the existing design codes in predicting the capacity of CFST columns. Feature importance analyses and SHapley Additive explanations (SHAP) explain the ML model performances and make the developed models interpretable. Resistance factor is determined using the best performing ML model for compressive strength prediction of CFST stub columns following AISC 360-16 code provision.