Explainable artificial intelligence framework for plastic hinge length prediction of flexural-dominated steel-reinforced concrete composite shear walls

Abstract

Due to the complex and nonlinear characteristics of the plastic hinge mechanism in steel-reinforced concrete composite shear walls (SRCCSWs), there are significant challenges in accurately predicting the plastic hinge length and providing a rational explanation for the mechanism. Advanced machine learning (ML) algorithms present a promising approach to tackle these issues. However, the development and interpretation of data-driven models for plastic hinge behavior are impeded due to the lack of a reliable and comprehensive analytical framework, as well as the inherent 'black box' nature associated with ML models. This study aims to establish an interpretable artificial intelligence framework for predicting the plastic hinge length of SRCSCWs. A reasonable understanding of the potential plastic hinge mechanism is provided. An experimental database encompassing two types of section reinforcement in SRCCSWs was compiled from existing literature. A Conditional Table Generative Adversarial Network (CTGAN) was employed to augment the training dataset and create a more extensive database, thereby addressing limitations posed by scarce available data. Eight individual and integrated ML algorithms were utilized to predict the plastic hinge length of SRCCSWs, and their performance and efficiency were comprehensively evaluated. The SHapley Additive exPlans (SHAP) method was introduced to systematically elucidate how design parameters affect plastic hinge mechanisms through feature importance analysis, feature dependency and interaction evaluation, and individual prediction interpretation. A simplified optimization equation is proposed by Bayesian methods to predict the equivalent plastic hinge length of flexural-dominated SRCCSWs based on the empirical models and this comprehensive database. Results show that the XGBoost model can make accurate and reliable predictions for plastic hinge length. Global interpretation reveals that geometric information makes a significant contribution to plastic hinge length. There is considerable interaction between the shear span ratio and the boundary element steel volume ratio. In addition, the optimization equation has shown commendable prediction accuracy for both synthetic and experimental data, although slightly lower than the accuracy of the XGBoost model, which can meet the requirements of engineering applications within a certain range.