Optimized machine learning models for prediction of effective stiffness of rectangular reinforced concrete column sections

Abstract

To evaluate the response of a reinforce concrete (RC) building subjected to nonlinear dynamic loading, it is recommended to use effective stiffness of the members. Currently, design codes and practices provide an empirical fraction of un-cracked stiffness to consider the effect of reduced stiffness in cracked sections. However, it is imperative to conduct an accurate assessment of effective stiffness as it directly affects the distribution of forces and various demands in nonlinear dynamic analyses. Manual calculation of effective stiffness based on strength is a time-consuming process. Therefore, the present study attempts to estimate the effective stiffness ratio of rectangular column sections using machine learning algorithms with ensemble learners namely, Adaptive Boosting (AdaBoost), Random Forest (RF), Extreme Gradient Boosting (XGBoost), Decision Tree (DT) and Categorical Gradient Boosting (CatBoost) using a database of 226 samples of nonlinear dynamic analysis results, obtained from previous work of the author. Additionally, Bayesian Optimization has been used for optimal hyperparameter selection. The results of the present study exhibit that the CatBoost model outperforms other considered machine learning models with R2 value of 0.9921 and 0.9966. A sensitivity analysis based on SHAP has also been implemented to establish the correlation between input parameters and output parameters.