Data-driven shear strength prediction of steel reinforced concrete composite shear wall

Abstract

Due to composition of concrete and steel materials, steel reinforced concrete composite shear wall (SRCCSW) plays a pivotal role as the primary lateral load-resisting component in buildings. Accurate prediction of shear strength is crucial, but numerous complex factors affect the shear strength, including shear-span ratio, axial load ratio, and materials properties. Therefore, it is imperative and urgent to develop precise prediction model. In this paper, machine learning (ML) models for predicting the shear strength of SRCCSW are developed. Firstly, a comprehensive database is established by collecting experimental data of 149 SRCCSW. Twelve predictive models are proposed, evaluated, and compared, including linear regression, decision tree, K-nearest neighbors, support vector regression, random forest, gradient boosting, adaptive boosting, categorical boosting, extreme gradient boosting (XGB), light gradient boosting machine, histogram-based gradient boosting, and artificial neural network. Then, six existing empirical models are further evaluated and compared. Finally, SHapley Additive exPlanations (SHAP) is employed to comprehensively explore the global explanation, individual explanation, and feature dependency of XGB model. The results show that the LR model is not as effective as that of other ML models, and the predictive performances of other ML models are generally comparable. The XGB model outperforms the best empirical model with a correlation coefficient of 0.99 (vs. 0.80) and a standard deviation of 520 kN (vs. 491 kN), while the experimental results have a standard deviation of 516 kN. Compared to other models, the XGB model possesses superior predictive accuracy and smallest dispersion. Wall height and shear-span ratio are the most two critical variables influencing the shear strength. At the individual level, the model’s prediction may depend only on certain features rather than having complex dependencies on all features. For the same feature, SHAP values may vary significantly due to the influence of other features. These results provided by SHAP reveal the detailed insights into the comprehensive impact of features on the shear strength of SRCCSW.