Machine Learning models to predict nonlinear simulations of net section resistance in steel structures

Resumo

This paper applies Machine Learning techniques to predict computer simulations of net section
resistance in bolted cold-formed steel connections of steel structures. Support Vector Regression (SVR) and
Gaussian Process Regression (GPR) techniques where chosen to be used for the analysis due to their good
performances on high dimensional data problems. One of the goals is to construct an efficient machine learning
model with minimal training to the uncertainty quantification of the net-section resistance. The algorithms were
trained using data set from expensive nonlinear finite element simulations where the resistance depends on the
cross-section geometry, connection eccentricity and connection length. The finite element simulations were
considered nonlinear due to the elastoplastic behavior of the steel. SVR and GPR were then compared by using
standardized statistics measures with different cross-validation strategies. The results showed that SVR had a
slightly better performance. In addition to that, it was possible to identify the best covariance function of each
technique for this specific problem.