Enhancing Circular CFST Columns Under Axial Load Compressive Strength Prediction and Inverse Design Using a Machine Learning Approach

This study presents a machine learning framework for predicting the axial compressive strength of circular concrete-filled steel tube (CFST) columns subjected to concentric and eccentrically applied axial loads. A harmonized database of 1287 test specimens was compiled, encompassing diverse material strengths, geometric configurations, and eccentricity levels. Among the trained models, the CatBoost (CatB) algorithm exhibited the highest predictive performance. A 300-run Monte Carlo simulation yielded a mean R2 of 0.966 (Min: 0.804; Max: 0.996), with a mean RMSE of 588.8 kN and MAPE of 8.36%, demonstrating accuracy and robustness across repeated randomized splits. Comparative benchmarking against current design equations revealed that CatBoost substantially reduced prediction scatter, improving the mean ratio and reducing the COV from 70–75% (ACI/AIJ/Wang) to 5.43%, while maintaining a nearly unbiased mean prediction ratio of 1.00. In addition, inverse prediction models based on CatBoost achieved test-set R2 values of 0.908 for compressive strength and 0.945, 0.900, and 0.816 for key design parameters (D, t, L), indicating promising capability for supporting preliminary sizing and parameter selection. The outcomes of this study highlight the potential of data-driven modelling to complement existing design provisions and assist engineers in early-stage decision-making for axially loaded circular CFST columns.