Interpretable Ensemble Machine Learning Models for Prediction of Shear Strength of Concrete Beams Reinforced with FRP Rebar

Abstract

A lack of design guidelines in Indian Standard code for fiber reinforced polymer (FRP) bars could result in over-reinforcement, driving up construction costs. The current design guidelines for shear strength vary based on several parameters and often provide conservative results. To address this issue, a machine learning (ML) framework based on ensembled models has been developed to estimate the shear capacity of FRP reinforced concrete (FRP-RC) beams. A well-curated dataset consisting of RC beams with FRP rebars without stirrups has been utilized for this purpose. An ensemble algorithm, XGBoost, and a traditional algorithm, Support Vector Regressor (SVR), have been compared for evaluating the shear capacity. The algorithms’ hyperparameters were optimized using GridSearch on the training set, combined with a ten-fold cross-validation optimization method. The models’ performance has been assessed using four metrics: R² score, RMSE, MAE, and MAPE. The best performing XGBoost model has achieved performance metrics of 0.94, 34.39 kN, 14.68 kN and 15.80% on testing dataset for R², RMSE, MAE and MAPE respectively. The model’s reliability has been further confirmed by comparing it with the current design codes ACI 440-1.R15 and GB50608-2020. Since ML-based models function as black boxes, a unified SHapley Additive exPlanations (SHAP) approach has been applied to interpret the outcome of XGBoost model. The beam width (b), span-to-depth (a/d) and effective depth (h) have been identified as the most significant input features which can improve the shear strength predictability for FRP-RC beams.