A machine learning based method for predicting the shear strength of Fiber-Reinforced Concrete joints in precast segmental bridges

Abstract

The utilization of Fiber-reinforced concrete (FRC) in the construction of Precast segmental concrete bridges (PSCBs) is progressively increasing. The design of joints among precast members can considerably impact the overall integrity and safety of the structure. Consequently, accurately estimating the shear capacities of FRC joints has become critical for evaluating the safety of PSCBs structures. However, the effectiveness and efficiency of formulas or models intended for prediction of the shear capacities of joints with normal concrete are unknown in predicting the shear capacities of FRC joints. In this study, the machine learning (ML) algorithms are employed to predict the ultimate load of FRC joints, utilizing a database of 192 push-off test results. Three individual models (Lasso, SVM, and KNN) and three ensemble models (RF, GBDT, and XGBoost) were developed for estimating the ultimate load. The GridSearchCV technique was utilized for parameter optimization based on the training set, in conjunction with a five-fold cross-validation method. A comparative study revealed that the XGBoost ensemble model surpassed other models, with a predictive performance ranking as follows: XGBoost > GBDT > RF > KNN > SVM > Lasso. The best-performing XGBoost model achieved performance metrics of R2 = 0.964, MAE = 41.318 kN, RMSE = 56.760 kN, and MAPE = 14.228%. The superior reliability of the XGBoost ensemble model was further validated through comparisons with current design codes and existing models. As ML-based models are typically black-box models, the SHapley Additive exPlanations (SHAP) method was employed to explicitly link the predicted output to the inputs. Based on SHAP results, confining stress (CS), depth of shear key (D), compressive strength of concrete (fcu), total height of joints (H), and height of flat part (Hsm) recognized as the crucial parameters affecting ultimate load.