Mechanics-informed neural network for predicting the axial compressive strength of short circular concrete-filled steel tube columns

Abstract

This study proposes a mechanics-informed neural network (MINN) model for predicting the axial compressive strength of short circular concrete-filled steel tubular (CCFST) columns by incorporating mechanics-based knowledge describing confinement effects, composite action and established mechanical relationships into the learning process. An experimental database comprising 1,071 short CCFST column specimens was collected from previously published studies and randomly divided into training and testing datasets at proportions of 75% and 25%, respectively. K-fold cross-validation was employed during model development to assess robustness and reduce dependence on a single training–validation split. The proposed MINN demonstrated excellent predictive performance, achieving a testing R2 of 99.47% and a mean predicted-to-experimental strength ratio of 1.051. To evaluate the interpretability of the developed model, SHAP, sensitivity and elasticity analyses were conducted. The results indicated that incorporating mechanics-based knowledge improved the mechanical interpretability of the model and enabled the MINN to capture mechanically meaningful relationships between the input variables and axial compressive strength. Furthermore, the analyses identified the outer diameter as the most influential parameter, followed by the steel tube thickness and material strengths, whereas the length-to-diameter ratio exhibited a comparatively minor influence, consistent with the expected behavior of short CCFST columns.