FluxGAN: A Physics-Aware Generative Adversarial Network Model for Discovery of Microstructures That Maintain Target Heat Flux*

In recent years, generative adversarial networks (GANs) are increasingly being used to predict multiphysics phenomena where the underlying physical model is complex or unknown. Physics-informed generative adversarial networks have been proposed for solving forward, inverse, and mixed stochastic problems using data from a limited number of scattered measurements. In parallel, GANs have shown remarkable progress for creating high-resolution synthetic images, supported by revolutionary advents in the graphical data domain. Until recently, training such model required enormous amount of training images and was highly unstable. However, recent developments enabled GANs to generate realistic microstructural data that can augment limited experimental data or simulate microstructures that are challenging to obtain experimentally. We propose a new physics-aware FluxGAN model that simultaneously generates high-quality microstructural images and high-resolution description of thermal properties. During the training phase, the model learns about the relationship between the structural features and the heat flux created due to external temperature gradients. Once trained, the model generates new structural and associated heat flux environments, bypassing the computationally expensive modeling. We demonstrate that the model can generate designs of thermal sprayed coatings that satisfies target thermal properties. The FluxGAN model uses synthesis-by-part approach and generates arbitrary large size images at low computational cost. The model is capable of generating coating microstructures and physical processes in three-dimensional domain after being trained on two-dimensional examples. Our approach has the potential to transform the design and optimization of microstructures for various applications, including coatings for high-temperature and long-duration operation of gas turbines for aircraft or ground-based power generators.

Ref: A. K. Pimachev, M. Settipalli, and S. Neogi. arXiv preprint arXiv:2310.04622 (2023).