Machine learning (ML) based models for predicting the ultimate bending moment resistance of high strength steel welded I-section beam under bending.

This paper presents an in-depth numerical investigation into the local buckling behaviour and develops advanced machine learning (ML) -based design methods for high strength steel (HSS) welded I-section beams for predicting the ultimate bending moment resistance under bending. Though HSS welded I-sections are commonly used in construction industry, the interaction effect between the flange and web plates as well as the relatively simplified design formulae result in limited accuracy of ultimate bending moment prediction and complex endeavour is involved for sections subject to local buckling. Finite element models are firstly developed and validated against the collected experimental test data, after which an extensive parametric study is carried out covering a larger range of cross-section slenderness and steel grades. Five ML models including Linear regressor (LR), Support vector regressor (SVR), Artificial neural network (ANN), Random forest regressor (RFR) and Boosting algorithm (XGBoost) were subsequently developed based on the test dataset (experimental and numerical data) to train and test the ML models. Though the current design codes of EN 1993-1-12 and AISC 360-16 can generally provide accurate cross-section classifications with appropriate slenderness limits, the ultimate bending moment resistance predictions are overconservative, particularly for slender sections. The developed ML models outperformed the codified design methods with notable improvements and can be employed in predicting the ultimate bending moment resistance of HSS welded I-section beams under bending. • Finite element modelling of high strength steel I-section is presented. • An extensive parametric study comprising 548 numerical models is carried out. • Slenderness limits specified in existing design codes are assessed. • Five machine learning models for predicting ultimate bending moment resistance is developed. • The performance of the developed machine learning models is evaluated. [ABSTRACT FROM AUTHOR]