Assessment of punching shear strength of FRP-RC slab-column connections using machine learning algorithms

Recently, the use of fiber-reinforced polymer (FRP) bars replacing steel reinforcement has been widely applied to overcome the corrosion issue, particularly concrete slab-column connections using FRP bars as flexural reinforcement (FRP-RC slabs). However, experimental studies showed that the use of FRP differentiates the punching shear behavior from steel-RC slabs. Various methods have been proposed to predict the punching shear strength of FRP-RC slabs, but existing design equations need improvement because their accuracy is low with wide scatteredness. Thus, this study aims at pioneering the application of machine learning (ML) algorithms for the prediction of the punching shear strength of FRP-C slabs without shear reinforcement. For this purpose, an experimental database with 104 specimens was compiled, with the input variables of the shear span-to-effective depth ratio, column perimeter-to-effective depth ratio, effective slab depth, concrete compressive strength, FRP reinforcement ratio, and ultimate tensile strength and elastic modulus of FRP. Three ML algorithms, including support vector regression (SVR), random forest (RF), and extreme gradient boosting (XGBoost), were evaluated for the application. To develop the ML-based models, a grid search method with a 5-fold cross-validation approach was used in the training process to determine the optimal hyperparameters. The performance of the ML-based models was estimated using various statistical estimators and compared with the current design codes and existing models. The comparisons showed that all three ML-based models could accurately predict the punching shear strength of the FRP-RC slabs without any significant and evident bias with the input variables. The XGBoost-based model displayed the best prediction with the coefficient of determination (R²) of 0.962, the root mean square error (RMSE) of 0.061 MN, mean absolute error (MAE) of 0.035 MN, and mean absolute percent error (MAPE) of 8.931% for testing dataset. The correlation coefficient, feature score, and sensitive analysis for the input variables indicated that the effective slab depth has the most substantial influence on the prediction performance. The prediction by the XGBoost-based model was more accurate and robust than that by the SVR- and RF-based models, current design codes, and existing models. These analysis results proved that the XGboost-based model can be used in the design and evaluation of FRP-RC slabs reliably and precisely.