Machine learning models for predicting concrete beams shear strength externally bonded with FRP

This paper focuses on preparing a database that covers a wide range of experimental data encompassing all the variations in concrete design and fiber reinforced polymer (FRP) composition in reinforced concrete (RC) beams. Artificial intelligence technique is applied to develop 10 machine learning models for estimating the shear ca-pacity of RC beams strengthened in shear using FRP. This study is unique in the sense that it covers the largest data regarding T-beams available in literature so far and is not limited to a particular beam cross section or FRP composition. A comparison among several design guidelines is done to evaluate the efficiency of the guidelines as well as the models developed in shear capacity estimation of FRP retrofitted RC beams. The results reveal that the ensemble learning models developed in this paper namely, random forest, CatBoost and XGBoost give the highest precision in estimating shear capacity of the FRP retrofitted RC beams. For rectangular beams the R2 and MAE values from CatBoost are 0.871 and 0.214 kN, respectively whereas those for XGBoost are 0.870 and 0.202 kN, respectively. The R2 and MAE values for T-beams using random forest are 0.897 and 0.128 kN, respectively and those using CatBoost are found to be 0.899 and 0.127 kN, respectively. The most significant input feature by Shapley Additive exPlanations is found to be the effective depth for the rectangular beams whereas height of FRP layers in the case of T-beams. The proposed ensemble models in this paper are proved to be superior to the existing mechanics-driven models currently being used for design practices.