Analysis of Steel Structural Profiles under Flexural-Compression stress using Machine Learning Algorithms

Resumo

The design of steel profiles requires a rigorous analytical approach, especially for elements subjected to flexural-compression, which involves the simultaneous action of compressive forces and bending moments. This type of loading is more complex than cases such as pure tension, pure compression, or flexural-tension, as it demands the combined consideration of both instability and interaction between acting forces. Given this complexity, conventional verification methods may be insufficient for rapid generalizations, particularly when analyzing multiple types of cross-sections. In this context, Machine Learning (ML) emerges as a promising alternative tool, capable of modeling complex relationships between input and output variables based on historical data. ML is a subfield of artificial intelligence that enables the development of computational models that learn from experience, recognizing patterns in data and making predictions without explicit programmed rules. This study proposes the application of machine learning algorithms to verify members under flexural compression stress. The algorithms K-Nearest Neighbors (KNN), Support Vector Machine (SVM), and Decision Tree were employed, combined with ensemble techniques and class balancing strategies to mitigate bias and variance issues. The implementation was conducted in Python using the Scikit-Learn library. The dataset was generated through a FORTRAN program, based on the NBR 8800:2024 standard criteria. The study considered multiple commercial steel profiles, common structural load combinations, different steel grades, and varied restraints along both principal axes. The diversity of these variables resulted in a high number of combinations, necessitating a robust computational tool to perform all calculations —a role fulfilled effectively by FORTRAN. The input features included section geometry, applied load, element length, steel class, and boundary conditions, while the output was a binary classification (OK/NOT OK) indicating structural compliance. The dataset was split into 80% for training and 20% for testing. The results showed satisfactory performance for all applied algorithms, with Random Forest – an ensemble method based on decision trees – standing out due to its high accuracy in estimating the class. This reinforces the potential of machine learning as a complementary tool to traditional approaches, especially in situations involving multiple variables.