Predicting the capacity of thin-walled beams at elevated temperature with machine learning
Carlos Couto, Qi Tong, Thomas Gernay
Abstract
For thin-walled beams in fire, the development of simple analytical models is hindered by the complexity of the interaction between local and lateral-torsional buckling combined with the temperature reduction of steel properties, resulting in overly conservative fire design rules. This paper investigates the application of machine learning models to predict the capacity of steel beams with thin-walled sections at elevated temperatures. Machine learning models provide a pathway to deliver fast and accurate methods to predict complex non-linear problems, which may overcome limitations of existing design methods, time-consuming finite element simulations, and expensive laboratory tests; yet these models remain unexplored in this particular field of study. This work describes the development, validation, and application of artificial neural networks, support vector machines, polynomial regression and random forests using an extensive dataset of numerical results from previously validated finite element models. It is shown that these models can also predict the capacity for beams with loading and boundary conditions outside the training dataset range. Finally, the machine learning models are compared against existing design proposals to demonstrate the benefits of using these advanced techniques to calculate the capacity of thin-walled beams at elevated temperatures.