Boosting machines for predicting shear strength of CFS channels with staggered web perforations. (December 2021)

• Record Type: Journal Article
• Authors: Degtyarev, V.V.; Naser, M.Z.
• Is Part Of: Structures. Volume 34(2021)
• Page Start: 3391
• Page End: 3403
• Publication Date: 2021-12
• Subjects: Cold-formed steel (CFS) -- Perforated web -- Shear buckling -- Shear capacity -- Machine learning -- Boosting machines; Structural engineering -- Periodicals; 624.1
• DOI: 10.1016/j.istruc.2021.09.060
• Languages: English
• ISSNs: 2352-0124

Abstract

Abstract: Cold-formed steel (CFS) purlins and studs with staggered web perforations have been used in construction to improve the thermal efficiency of buildings. The perforations adversely affect structural properties of the members, especially their shear strength. Accurate shear strength predictions of such members is a challenging task, which is not easily solved by the traditional methods. This paper explores five machine learning (ML) boosting algorithms, including gradient boosting regressor (GBR), extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), gradient boosting with categorical features support (CatBoost), and adaptive boosting (AdaBoost), as tools for predicting the elastic shear buckling loads and the ultimate shear strength of CFS channels with staggered web perforations. The models were developed using open-source software and a large dataset of finite element simulation results with 3512 samples for each property. The models' optimal hyperparameters were established through an extensive tuning process to ensure their best performance measured using the robust ten-fold cross-validation method. The elastic buckling loads and the ultimate shear strengths predicted by the developed models compared excellently with the simulation results, with CatBoost being the most accurate model in predicting both properties. The models' accuracy exceeded the accuracy of the existing descriptive equations. The CatBoost models were inspected by evaluating the relative feature importance and partial dependence of the model-predicted targets from features, which aligned well with the mechanics-based knowledge and confirmed the ability of the models to capture the underlying physics. The predictions of the developed boosting machines were compared with those of previously developed ML models, which allowed for ranking the models and selecting the most accurate ones. Resistance and safety factors for the shear strength predicted by the ML models were determined using the AISI S100-16 w/S2-20 provisions. A Python-based interactive notebook was created for rapid and accurate predictions of the elastic shear buckling loads and the ultimate shear strengths of the slotted channels using the developed models.