Prediction of experimental test results of thin-walled racking components using machine learning

This thesis study aimed to predict the capacity of cold-formed racking components employing 4 types of machine learning (ML) algorithms. Within this scope, a database consisting of 240 distortional buckling tests and 30 floor connection tests has been used by covering different geometric features and loading conditions. Linear, ridge and support vector regression (SVR), and artificial neural network (ANN) methods have been compared in predicting the results. Geometric features of the cold-formed samples are selected as input parameters to train the ML algorithms. The target features were the maximum axial load and bending moment capacities, respectively for the distortional buckling and floor connection tests. The ANN was the best algorithm that succeeded to predict the test results of distortional buckling tests with minimum 0.82 R2 score and maximum 3.5% absolute error. The SVR was observed to be the best for the floor connection tests considering overall performance through three validation sets, resulting in a maximum 16.7% absolute error for the bending moment capacity prediction.