Existing semi-empirical formulas for predicting punching shear capacity in FRP bar reinforced concrete flat slabs without shear reinforcement often prove inaccurate and unstable. This is primarily due to limited modeling data, inadequate consideration of key variables and neglect of complex nonlinear relationships. To address these challenges, this study delves into the utilization of advanced machine learning (ML) algorithms to offer precise and dependable estimates of punching shear capacity in such structural components. The study initially compiled a comprehensive database comprising 165 sets of test data, integrating eight crucial variables for model development. Subsequently, four data-driven models including back propagation artificial neural network (BPANN), support vector regression (SVR), random forest (RF) and gradient boosting regression tree (GBRT) were formulated to estimate the shear capacity. The efficacy of these models was assessed in comparison to existing prediction formulas. To interpret the models, this study also introduced shapley additive explanation (SHAP) and partial dependence plot (PDP) to quantitatively evaluate the influence of variables on predicted results. Research findings suggest that: (a) Among 25 existing formulas, Ju et al.’s approach performs notably well, with R2, Pre/Exp, MAPE and RMSE values at 0.76, 1.02, 22.2 % and 142.8 kN, respectively. (b) ML models surpass traditional formulas in predictive accuracy, with R2, Pre/Exp, MAPE and RMSE values ranging from 0.89 to 0.93, 1.03–1.09, 4.8–9.5 % and 55.4–69.0 kN, respectively. The GBRT model demonstrates the highest precision. (c) SHAP analysis of the GBRT model reveals that effective slab height and column section aspect ratio are pivotal variables influencing punching shear capacity. (d) PDP analysis quantitatively illustrates how punching shear capacity varies with each key variable.
Full Text Finder