Capacity prediction and failure mode classification of cold-formed steel built-up columns using machine learning methods

• Five machine learning models were used to predict capacity of CFS built-up columns. • Five machine learning models were employed to classify the failure of CFS built-up. • MLP and XGBoost were the best ML methods to predict CFS built-up capacity. • ML classifiers have demonstrated accuracy in classifying the buckling modes of CFS built-up columns. Combining two or more cold-formed steel (CFS) sections via fasteners to make built-up columns (BCs) is an innovative way to increase the loading capacity of CFS structural framing systems. Predicting the capacity of CFS-BCs can be challenging due to the variety of available cross-section geometries and the interaction of the connected columns. Current analytical and mechanical models in the literature have limitations in predicting the capacity and failure modes of CFS-BCs. Therefore, considering the limitations of conventional methods, five machine learning methods were trained and tested using an open-access CFS-BCs database compiled by the authors containing a data set of 1037 specimens to predict the axial capacity and classify the buckling failure modes of CFS-BCs. These techniques included Random Forest (RF), Extreme Gradient Boosting (XGB), Support Vector Machines (SVM), K-Nearest Neighbours (KNN), and Multilayer Perceptron (MLP) of artificial neural networks. Regression models were compared with the Direct Strength Method (DSM) in the North American Specification for the Design of Cold-Formed Steel Structural Members. The best regression and classification models were interpreted using the Shapley Additive Explanations (SHAP) and the Local Interpretable Model-Agnostic Explanations (LIME), respectively. The models' performance metrics indicated that MLP and XGB were the best in predicting axial capacity, while RF excelled in classifying failure modes, including interactive buckling. MLP and XGB showed better accuracy than DSM. Interpretation results indicated that machine learning models recognised the underlying mechanics of designing compression CFS-BC members.