Hybrid machine learning with Bayesian optimization methods for prediction of patch load resistance of unstiffened plate girders

Abstract

This paper aims to propose a new hybrid Machine Learning (ML) with Bayesian Optimization (BO) methods for predicting the patch loading resistance, P_u of longitudinally unstiffened plate girders. A total of 354 tests of the unstiffened plate girder under patch loading are collected and used for the training and testing to establish the proposed models. Five ML models including Support Vector Machines (SVM), Decision Tree (DT), Gradient Boosted Tree (GBT), Extreme Gradient Boosting algorithm (XGBoost), and CatBoost regression (CAT) are employed, and the BO method is used to optimize the hyperparameters of these ML models, in order to show which of them is best-suited for prediction of the PLR of longitudinally unstiffened plate girders. It was found that the BO-GBT presents the best accuracy compared to others. The performance of the BO-GBT model is validated by comparing its predictive results with the current design standards and the existing formulae. Additionally, the Shapley Additive Explanations (SHAP) method is employed to evaluate the importance and contributions of each input variable on the proposed model, and a Graphical User Interface (GUI) tool is developed to conveniently estimate the P_u of the unstiffened plate girders. Finally, the BO-GBT model is used to develop a support tool for finding suitable geometric dimensions and material properties of longitudinally unstiffened girder under patch loading in the preliminary design stage. The optimization tool is accessible online for the users for more convenient use in practical design purposes.