Development of deep learning-based joint elements for thin-walled beam structures

摘要

This study presents a new modeling technique to estimate the stiffness matrix of a thin-walled beam-joint structure using deep learning. When thin-walled beams meet at joints, significant sectional deformations occur, such as warping and distortion. These deformations should be considered in the one-dimensional beam analysis, but it is difficult to explicitly express the coupling relationships between the beams' deformations connected at the joint. This study constructed a deep learning-based joint model to predict the stiffness matrix of a higher-order one-dimensional super element that presents the relationships. Our proposition trains the neural network using the eigenvalues and eigenvectors of the joint's reduced stiffness matrix to satisfy the correct number of zero-strain energy modes overcoming the randomly perturbed error of the deep learning. The deep learning-based joint model produced compliance errors mostly within 2% for a given structural system and the maximum error of 4% in the worst case. The newly proposed methodology is expected to be widely applicable to structural problems requiring the stiffness of a reduction model.