Multiphysics Design Optimization via Generative Adversarial Networks

This work presents a method for generating concept designs for coupled multiphysics problems by employing generative adversarial networks (GANs). Since the optimal designs of multiphysics problems often contain a combination of features that can be found in the single-physics solutions, we investigate the feasibility of learning the optimal design from the single-physics solutions, to produce concept designs for problems that are governed by a combination of these single physics. We employ GANs to produce optimal topologies similar to the results of level set topology optimization (LSTO) by finding a mapping between the sensitivity fields of specific boundary conditions, and the optimal topologies. To find this mapping, we perform image-to-image translation GAN training with a combination of structural, heat conduction, and a relatively smaller number of coupled structural and heat conduction data. We observe that the predicted topologies using GAN for coupled multiphysics problems are very similar to those generated by level set topology optimization, which can then be used as the concept designs for further detailed design. We show that using a combination of multiple single-physics data in the training improves the prediction of GAN for multiphysics problems. We provide several examples to demonstrate this.