Machine learning-based probabilistic predictions for Concrete Filled Steel Tube (CFST) column axial capacity

Abstract

This study presents a novel probabilistic machine learning (ML) approach using Natural Gradient Boosting (NGBoost) to predict the axial compressive capacity of Concrete Filled Steel Tube (CFST) columns. Leveraging a comprehensive dataset of 1,127 experimentally tested CFST specimens under axial compressive loads, we compare the performance of various ML algorithms. These include deterministic models like eXtreme Gradient Boosting (XGBoost) and Artificial Neural Networks (ANN), and probabilistic models such as XGBoost-Distribution (XGBD) and NGBoost. The NGBoost model, which employs Normal and LogNormal distributions to account for uncertainties in input data, demonstrates superior predictive accuracy and robustness. SHapley Additive exPlanations (SHAP) are utilized to interpret the influence of input features, providing insights into the relative importance of different structural parameters. The predictive performance of the NGBoost model with LogNormal distribution is benchmarked against existing design codes, including Eurocode 4, ANSI/AISC 360-22 AS/NZS 2327, and Chinese Standard (GB50936-2014), showcasing its enhanced accuracy and reliability. This approach not only improves predictive performances but also integrates uncertainty quantification, making it highly suitable for design applications in Civil Engineering where understanding the variability in the structural behavior is crucial. • Probabilistic ML model ( NBGoost ) predicts CFST column capacity with high accuracy. • Comprehensive dataset of 1,127 CFST specimens used for model training and validation. • SHAP analysis provides insight into key features influencing CFST column capacity. • NGBoost outperforms existing design codes like Eurocode 4, ANSI/AISC 360-22 and GB50936-2014. • Uncertainty quantification integrated, enhancing reliability for Civil Engineering design.