TOPOLOgyGAN: OPTIMIZATION OF TOPOLOgy USING GENERATIVE ADVANCED NETWORKS BASED ON PHYSICAL FIELDS OVER THE INITIAL DOMAIN

Abstract

Load and boundary constraints expressed as vectors or sparse matrices in deep learning topology 
optimization often lack the chance to incorporate a full understanding of the design challenge, resulting 
in less than optimal generalization outcomes. We present TopologyGAN, a novel data-driven topology 
optimization approach that uses different physical fields calculated on the initial, unoptimized material 
domain as inputs to the generator of a conditional generative adversarial network (cGAN). On test issues 
with previously unknown boundary conditions, TopologyGAN achieves a roughly 3× decrease in the 
mean squared error and a 2.5× reduction in the mean absolute error when compared to a baseline cGAN. 
We also provide a hybrid network for the generator dubbed U-SE (Squeeze-and-Excitation)-ResNet, 
which is based on a number of current network models and significantly improves overall accuracy. We 
make our trained network and complete implementation available to the public.