Improved machine learning modeling for predicting shear capacity of RC deep beams

Abstract. Pile caps, underground structures, bridges, floor diaphragms, and multistory RC structures use RC deep beams (RCDBs), the fundamental structural components carrying heavy loads. RCDBs have a lower clear span-to-total depth ratio (usually below 4) with distinct structural behaviour like slender beams. Despite being widely used, RCDBs shear design is still challenging due to the applicability and reliability of different design techniques considering the architectural configuration/restriction, code provisions, and other nonlinear variables for optimal geometric design. RCDBs are prone to shear failures, initiating an abrupt, catastrophic collapse that poses severe challenges to public and infrastructural assets. Previous studies developed machine learning (ML) models for forecasting RCDBs’ shear strength, overlooking some influential features and yielding low validation accuracy compared to experimental work. This paper deploys data-driven techniques to forecast the shear strength of RCDBs using more reliable results and accuracy than previous research studies. This study employed two different ML models, including Multi Expression Programming (MEP) and eXtreme Gradient Boosting (XGBoost), to forecast the shear strength of the RCDBs. The established database collected from experimental studies consists of 224 data points with 15 key input features, which determine the shear mechanism of the RCDBs. The dataset was split using the K-fold method into thirty percent data points to test model and remaining seventy percent for training. Shapley explanation identified Top Plate Width (Wtp) from the experimental setup and nominal yield strength of steel (Fyk) from material properties as the most influential factors. The model's accuracy and applicability were evaluated using several statistical parameters, including correlation coefficients (R2), mean absolute errors (MAE), Root mean square error (RMSE), a-10 Index and Median Absolute Error (MdAE). With R2 of 0.979 for validation, MAE of 30.842, RMSE of 32.848, MdAE of 29.58296 and a-10 Index of 0.4478, respectively, the XGboost model surpasses all the previous studies. Compared to MEP, the developed XGBoost model performed well overall, exhibiting excellent prediction accuracy in testing and validation datasets that forecasted the RCDBs’ shear strength . However, to make the model more practical in the design industry, this study proposes the collection of datasets with different concrete mixture compositions and reinforcement profiles. Moreover, applying Physics Informed Neural Networks (PINNs) on a more extensive and diverse dataset is advisable to decrease overfitting and in situ consideration of RCDBs.