LAUSR

Abstract Connections are a key component of steel structures. Current methods of analysis of steel frames mostly rely on approximate joint models based on zero-length rotational springs or mechanical models of rigid bars and springs. A method is presented in this paper to accurately model the stiffness of 2D welded steel joints that is based on surrogate modelling. The joint modelling process starts by developing an accurate 4-node (12 degree of freedom) cruciform finite element from a complete finite element model using substructuring techniques. Then a modal decomposition of these matrices yields the eigenvalues and mode shapes that are used to develop the design points of the training and validation sets of the surrogate model. The modal decomposition leads to a reduced set of variables that accurately models the design space. A uniform Rectangular Grid of design sites is proposed. The Kriging method and Support Vector Machine Regression are used to generate the metamodels and predict the stiffness at untried sites. Both methods provide accurate results. Simulations of steel frames are performed to show the accuracy of the proposed surrogate modelling.