General structure representation for neural networks

Current applications of neural networks for structural analysis are limited to specific structures. In order to expand the use of neural networks to general structural types, the input vector has to be capable of encoding all necessary aspects needed for the analysis. Furthermore, this vector has to be capable of representing any structure. A neural network trained with this general representation will not be limited to specific structures, but to any structures that can be fully modeled by the encoding technique. The structural stiffness matrix is proposed as input vector in this study. Several general numerical representations for structures were investigated. A neural network was trained with various examples using the global stiffness matrix as input and displacements as target output. The results of testing the trained network with stiffness matrices representing different structures are presented.