Enhanced data-driven shear strength predictive modeling framework for RCDBs using explainable boosting-based ensemble learning algorithms coupled with Bayesian optimization
Imad Shakir Abbood, Noorhazlinda Abd Rahman, B.H. Abu Bakar
Abstract
Despite over 70 years of investigation into the behavior of reinforced concrete deep beams (RCDBs), it remains challenging to accurately predict their shear strength (SS) due to the underlying intricate mechanism. Nowadays, there is a boom in implementing machine learning (ML) approaches in solving various structural engineering problems. This research aims to present a novel data-driven predictive framework for the SS of RCDBs using an explainable ML approach and incorporating a large database compilation of 950 experimental specimens with different web reinforcement. To achieve this goal, four ensemble boosting ML algorithms, namely Gradient Boosting, HistGBoost, XGBoost, and LightGBM, were adopted for implementation. The theoretical background and fundamentals were presented in detail. The predictive framework was systematically developed and enhanced through a customized procedure involving features selecting, data preprocessing, hyperparameter tuning, as well as model evaluating and explaining. Additionally, the predictive framework was compared with existing conventional shear-mechanism models to demonstrate its superiority. The results indicated that the proposed framework has excellent performance and performed better than conventional models, and the XGBoost model achieved the highest prediction accuracy and lowest errors among all models. Furthermore, the explainability analysis indicated that section width, concrete compressive strength, section height, and longitudinal reinforcement ratio have the largest positive impact on SS prediction of RCDBs, while shear span-to-effective depth ( ) and effective span-to-height ( ) have the largest negative impact. The provided findings are significant for design practice and deliver essential information for designers. Consequently, the proposed data-driven framework could enhance the understanding of the SS of RCDBs and offer guidance for practical applications.