Exploring machine learning for predicting elastic buckling and ultimate moments of steel decks in bending

This paper explores the use of machine learning in the form of the Support Vector Machine regressor (SVR) for predicting elastic buckling and ultimate moments of steel decks. SVR is a supervised learning algorithm chosen in this study due to its abilities for good generalization, nonlinear data handling, and efficient training on relatively small datasets.

The dataset for training and testing of SVR models consisted of finite element (FE) simulation results on steel decks commonly used in North America. The FE shell models of steel decks were developed in ANSYS and validated on available test data. Previously reported data on the elastic buckling and ultimate moments of steel deck were complemented by 960 results of simulations conducted in this study, resulting in a total of 1408 dataset samples. The following parameters were varied in the FE models: deck type, steel thickness and yield stress, intermediate longitudinal flange stiffener height and angle, deck span length, and bending orientation.

Eight SVR models were developed, validated, and tested. The models allow for predicting the following deck properties: critical elastic buckling moments corresponding to the local buckling of stiffened and unstiffened flanges, the distortional buckling of the webs (along with unstiffened flanges connected to the webs) and the flanges with intermediate longitudinal stiffeners; ultimate moments; and plate buckling coefficients of stiffened and unstiffened flanges, as well as flanges with intermediate longitudinal stiffeners. The developed models demonstrated good generalization and excellent prediction accuracy, which exceeded the accuracy of the AISI S100 provisions.

Learning Objectives:
Describe the process of applying machine learning techniques to structural engineering problems.

View content