Data-driven design approach for the lateral-distortional buckling in steel-concrete composite cellular beams using machine learning models

Abstract

Composite steel-concrete beams are presented as an interesting solution for covering long spans. These profiles are characterized by the existence of sequentially spaced openings along the beam web. A very common particular use of this type of structure is in continuous or semi-continuous floors, in which the existence of intermediate supports in the span of the structural member leads to the occurrence of points with negative bending moment. Steel structures in this condition can present different types of instability due to the compression of the bottom flange, such as lateral-distortional buckling and web-post buckling. In this context, a comprehensive parametric study was carried out through finite element simulation aiming to obtain the resistance capacity of this type of structural system. It was observed that different analytical procedures aimed at calculating the ultimate moment of this type of structure fail to predict the bearing capacity of the elements successfully. Thus, aiming to promote more precise design approaches, the modeling of the resistant load of these structures was performed using different supervised machine learning techniques. It was found that machine learning techniques have the potential to correctly predict the bearing capacity of the studied members, having a better agreement with the numerical data obtained, when compared to the approaches presented by analytical procedures described in the literature and design codes. Different machine learning algorithms were trained based on a dataset obtained via FEM: artificial neural networks, support vector machines, XGBoost and Random Forest. All models showed considerable effectiveness, with statistical determination coefficients (R²) higher than 0.9. Among these, the best prediction capacity was obtained with the XGBoost model in terms of correlation coefficient and Mean Square Error, presenting itself as the model that came closest to the set of numerical data used as reference.