Unified machine-learning-based design method for normal and high strength steel I-section beam-columns

High strength steel is regarded as a promising construction material due to its superior mechanical properties. However, the codified failure load predictions for high strength steel welded I-section beam-columns are not accurate in some cases owing to the lack of relevant design codes for high strength steel structures. In addition, current design codes ignore the interaction effect of the cross-section plate elements, leading to the inaccuracy of codified predictive ultimate resistances for both normal and high strength steel beam-columns. In this paper, a unified and accurate design method was proposed based on machine learning for both normal and high strength steel welded I-section beam-columns failing by different failure modes. Firstly, a total of 812 experimental and numerical data on welded I-section beam-columns with various steel grades, geometric dimensions, including cross-section dimensions and member lengths, and failure modes, including global buckling, local buckling and local-global interactive buckling were collected to establish a database. Seven machine learning algorithms, including Decision Tree, Random Forest, Support Vector Machine, K-Nearest Neighbour, Adaptive Boosting, Extreme Gradient Boosting and Categorical Boosting were applied to develop regression models to predict failure loads. Based on the established database, each machine learning model was then trained and key hyperparameters were optimised. The model performance was evaluated through a series of statistical indicators and the feature importance analysis, and the evaluated results indicated that the XG-trained regression model had the highest level of accuracy. Finally, the predictions given by the XG-trained regression model were compared with those of the existing codified design provisions, as given in Eurocode and American codes. The evaluation results indicated that the codified predictive bearing capacities of I-section beam-columns were scattered and inaccurate, while the XG-trained regression model provided substantially improved failure load predictions.