The research goal of this paper is to establish a machine learning (ML) model of shear capacity of reinforced concrete (RC) columns that can reflect the shear mechanism. The main work includes building database, training ML model, and interpretability analysis of optimal model. Aiming at the core problem of interpretability of ML methods, this paper proposes an interpretability method for the shear mechanism of RC columns. An experimental database including 427 sets of results of tests on RC columns is established, and eight salient input features are discerned through both mechanical mechanism assessment and Pearson correlation analysis. Twelve models, consisting of six basic ML models and six ensemble ML models, are verified against the database. The assessment reviews that the Light Gradient Boosting Machine (LightGBM) model was demonstrated to be superior to other models, achieving R2 values of 0.999 and 0.987 in training and test sets respectively. Additionally, compared to five semi-empirical models, the results from the LightGBM model also provides more accurate predictions and adeptly considered the influence of parameters. The Shapley additive explanation (SHAP) method and the explainable method based on the RC column shear mechanism proposed by this paper are used to explain the working mechanism of the LightGBM model. The proposed explanatory method of shear contribution rate is combine the additivity of SHAP value with the additivity of shear component contribution, the proposed explanatory method of shear contribution rate. An evaluation of SHAP feature Importance and SHAP feature dependence revealed that LightGBM aligns closely with the inherent shearing mechanism of the RC column. The shear contribution ratios of the concrete, stirrup, and axial loading in RC columns for the LightGBM model are similar to those predicted based on the semi-empirical models.
Read full abstract