Prediction of shear strength of RC deep beams using XGBoost regression with Bayesian optimization

For the shear strength of deep beams, most of current design codes adopt the strut-and-tie model. However, since the shear resistance mechanism of deep beams is quite complicated and shear failure could be catastrophic with no prior warning, more reliable and robust shear strength prediction models are required. To this end, in the present study, using the extreme gradient boosting (XGBoost) and Bayesian optimization (BO) algorithms, a BO-XGBoost hybrid model was developed and its hyperparameters were optimized based on a database of 320 test specimens. The prediction accuracy of the proposed BO-XGBoost hybrid model was compared with G-XGBoost (with hyperparameters tuned by grid search), R-XGBoost (with hyperparameters tuned by random search), D-XGBoost (with default hyperparameters), and existing design equations, and the comparison showed that the proposed model can predict most accurately the shear strength of deep beams without any significant and evident bias. The parametric study using the proposed model revealed that the effective depth, beam width, shear span-to-depth ratio, tension reinforcement ratio, and concrete compressive strength are critical to the shear strength of deep beams. For ease and convenience of the utilization of the proposed model, a stand-alone application program was also developed.