Machine learning models on the rigidity of welded shear tab connections

Abstract This article presents an analytical study on the rigidity of shear tab connections, building on a previous study that developed 281 finite element models of varying geometry and size. The initial study proposed an equation to predict shear tab connection behavior based on geometric characteristics. This article further refines those predictions by assessing uncertainties using supervised machine learning (ML) models and proposing improved equations and methods. First, a simplified polynomial regression curve with enhanced statistical metrics is introduced. Additionally, support vector machine (SVM) with feature dimension reduction is used to predict the rigidity of shear tab connections. To identify the best regression model, three algorithms were tested: random forest (RF), XGBoost (XGB), and artificial neural network (ANN), with XGB achieving the lowest error. For classification, four algorithms – SVM, ANN, RF, and CatBoost (CB) – were employed to categorize shear tab behavior as either simple or semirigid. Among these, CB demonstrated the highest accuracy in classifying new test data based on numerical modeling results.