Accurate axial-capacity prediction of concrete-filled steel tubular (CFST) columns is challenging because the response depends on section shape, material strength, tube proportion, slenderness, and load eccentricity. This study evaluates a mechanics-aware machine-learning workflow for circular and rectangular CFST columns using a compiled experimental database of 2276 specimens. The workflow uses a normalized response, \(\eta =P_{\textrm{exp}}/P_0\) , to reduce gross-size dominance while retaining a direct link with structural mechanics. Benchmark learners, a stacked ensemble, design-code predictors, SHAP interpretation, split conformal intervals, and a desktop deployment prototype were assessed under grouped validation and a limited temporal holdout. The stacked model achieved \(\hbox {RMSE}_\eta =0.129\) and MAPE \(=12.72\%\) under grouped cross-validation. On the 92-specimen temporal holdout, it achieved MAPE \(=8.80\%\) and placed 93.48% of predictions within \(\pm 20\%\) of the experimental capacity. However, the holdout was small, recent large-scale circular specimens remained difficult, and conformal intervals showed lower-than-nominal temporal coverage. The results suggest that mechanics-aware machine learning can complement conventional CFST capacity assessment, but further external validation and improved uncertainty calibration are needed before routine design-level use.
