Modeling the deformation capacity of reinforced concrete shear walls using artificial intelligence methods

Experimentally testing the behavior of reinforced concrete walls and modeling them using the finite element method involves significant limitations due to the high time and cost required. In recent years, the increase in open data and code sharing has led to a growing trend towards machine learning (ML) in applications. Within the scope of this study, the plastic rotation, reinforcement, and concrete unit deformation values at the section ends of 2548 wall models generated using the finite element method were examined using ML methods. First, plastic rotation was predicted using four tree-based methods (Decision Tree, Random Forest, Adaptive Boosting, Extreme Gradient Boosting) with various input parameters, and the performance of the models was evaluated considering the 10-fold cross-validation criterion. As a result of the comparisons, the XGBoost model, which showed the best performance, was retrained with hyperparameter tuning using Random Search and Optuna. The XGBoost-Optuna configuration was determined to yield the highest success. The results of this model were interpreted using SHAP analysis to reveal the relative effects of the variables and their possible interactions. This modeling approach, proven successful in plastic rotation prediction, yielded similarly effective results when applied to curtain unit deformations. To derive equations from the results, a sum of additive power law regression model was created separately for both actual (ANSYS) measurements and ML prediction outputs. The equation derived using the ML method was found to better represent the behavior. The findings show that the behavior of reinforced concrete walls can be predicted with high accuracy, speed, and interpretability using ML.