Prediction accuracy of Random Forest, XGBoost, LightGBM, and artificial neural network for shear resistance of post-installed anchors

Post-installed anchors and reinforcing bars are used to connect equipment or to fasten strengthening members to reinforced concrete (RC) structures. For safety reasons, appropriate structural design is critical. Recently, arti-ficial intelligence (AI) and machine learning (ML) have been applied in various fields. According to previous studies, the bending strength of the RC beam and the bond strength of the surface can be predicted using ML. In this study, the mechanical behavior of post-installed anchors subjected to shear force were predicted using ML. Four algorithms were applied in this study: Random Forest (RF), XGBoost (XB), LightGBM (LG), and an artificial neural network (ANN). Moreover, the authors' previous test results were used for the ML and testing. The number of specimens was thirty-two. The test parameters were the concrete compressive strength fc, diameter of the anchor bolt dd, type of adhesive, and tensile ratio rN. The values for fc and dd were set at 13.0-35.5 N/mm2 and 13-25 mm, respectively. In this study, one epoxy adhesive and three cement-based adhesives were used. rN, which is the ratio of the tensile stress to yield strength of the anchor bolt, was set to 0, 0.33, and 0.66. Conse-quently, the four algorithms could accurately predict the mechanical behavior of the specimen when the pa-rameters were within or close to the training data. However, the prediction agreements of RF, XB, and LG were not good for the behavior of specimens whose parameters were not included in the training data. Nevertheless, the ANN was able to reasonably predict the behavior of these cases. It was concluded that the four algorithms can make good predictions when the parameters are within or close to the training data. However, when parameters outside the training data were used, the ANN was the best of the four algorithms used in this study.