Machine learning based prediction model for plastic hinge length calculation of reinforced concrete structural walls

Abstract

Reinforced concrete structural walls (RCSWs) are integral part of buildings and other structures and are used to carry in-plane and out-plane loads. For assessment purposes and to ensure safety and resilience of the structure, the curvature, capacity, and strain demands of RCSWs need to be estimated. Nonlinear numerical models are increasingly used in earthquake engineering design and assessment, where it is critical to develop high fidelity simulation tools to precisely forecast the global and local behavior of RCSWs. Plastic deformations concentrated in small regions, i.e. plastic hinges, and characterized through the concept of plastic hinge length (PHL), are conventionally used to define the inelastic response of RCSWs. In this study machine learning (ML) algorithms were leveraged in evaluating the PHL of RCSWs. A database containing 721 planar and nonplanar RCSW samples were utilized for training and testing ML models. Four different algorithms were employed, namely: XGBoost, CatBoost, Random Forest (RF), and genetic programming. The RF model outperformed other ML counterparts. The results of the best ML model were also compared against empirical equations available in literature and nonlinear regression analysis. The proposed ML model provided much better predictions than the existing relationships. Predictions of most empirical models were found to be extremely over (or under) conservative. The proposed ML model was also presented via an online tool where users can obtain the PHL of any RCSW simply by inputting wall parameters and clicking a run button.