APPLICABILITY OF COORDINATE SYSTEM INDEPENDENT NEURAL NETWORKS ON COMPLICATED SYSTEMS

Abstract

 Coordinate system independence is required for a neural network of solving physical problems. We developed a synthesized post-processing for the full satisfaction of the coordinate system independent neural network. In this paper, we study the applicability of the synthesized post-processing to a practical neural network of analyzing structural seismic responses of a set of lumped mass spring models as a surrogate model for a nuclear reactor. It is shown that the neural network implemented with the synthesized post-processing can produce seismic responses that satisfy the coordinate system independence for all mass points on the complicated system. The increase of the neural network accuracy is discussed due to the satisfaction of coordinate system independence by using the synthesized post-processing. The increase of computational time due to the implementation of the synthesized post-processing is discussed by comparing with other coordinate system independent neural networks.