Implementing data-driven machine learning technique to estimate the shear strength of FRP-RC deep beams without stirrups

Abstract

The use of fiber-reinforced polymer (FRP) rebars as a substitute for steel rebars has introduced notable variations in the shear behavior of concrete members. To anticipate the shear strength of FRP-RC deep beams, numerous models, codes, standards, and guidelines have been established. The majority of existing shear design provisions for FRP-RC deep beams are empirically calculated or calibrated based on limited test results. The present study aims at developing a simple yet efficient shear strength prediction model for FRP-RC deep beams without stirrups using machine learning (ML) algorithm that does neither rely on crude assumptions nor require complicated calculations. Twelve ML models,including Linear Regression (LR), Ridge Regression (RR), Lasso Regression (LaR), Decision Tree (DT), K-Nearest Neighbour (KNN),Artificial Neural Networks (ANN),Categorical Boosting (CB), Adaptive Boosting (AB), Gradient Boosting (GB), Light Gradient Boosting (LGB), Extreme Gradient Boosting (XGB), and Random Forest (RF), were developed using a datasetof 245 data consisting of 161 experimental and 84 numerical resultsconsidering all key variables. The performance of the ML models was evaluated using various statistical measures and a comparison among the various design provisions was conducted to assess their effectiveness in shear capacity estimation of FRP-RC deep beams. Results revealed that the ensemble ML models exhibited better performance compared to the single ML models. The superiority of the ensemble models such asXGBoost (XGB), CatBoost (CB), and Random Forest (RF) models was confirmed with an accuracy of 92%, 91% and 90%, respectively significantly outperforming the current design practices and widely used empirical formulas. Among the twelve ML models, the XGBoost model is the most accurate model with the highest coefficient of determination (R^2) of 0.920 and least root mean square (RMSE), and mean absolute error (MAE) of 48.280, and 33.310 respectively.The model interpretation was performed through Feature Importance Analysis (FIA), SHapley Additive exPlanations (SHAP), and Individual Conditional Expectation (ICE) for the ensemble ML models to explain the model output compared with a black box. The Feature Importance Analysis (FIA) revealed that for the XGBoost model, beam depth (d), shear span-to-depth ratio (a/d), and beam width (b_w)were the most influential factors in predicting the shear strength, contributing 56.91%, 16.96%, and 12.84%, respectively. SHAP analysis and ICE plots demonstrated that beam width (b_w), depth (d), compressive strength (f_c^'), longitudinal reinforcement ratio (ρ_f), and modulus of elasticity (E_f) positively influenced the shear strength, while the shear span-to-depth ratio(a/d) had a negative impact.Additionally, the model performance was analyzed using Taylor diagram which further confirmed that the superiority of the XGBoost model. The proposed data-driven ML models demonstrated a high level of accuracy and excellent performance and were superior to the existing shear strength models. Finally, a simplifiedGraphical User Interface (GUI) was developed to aid practicing engineers when estimating shear strength without the need for complicated design procedures.