Virtual microstructure design for steels using generative adversarial networks

The prediction of macro-scale materials properties from microstructures, and vice versa, should be a key part in modeling quantitative microstructure-physical property relationships. It would be helpful if the microstructural input and output were in the form of visual images rather than parameterized descriptors. However, only a typical supervised learning technique would be insufficient to build up a model with real-image-output. A generative adversarial network (GAN) is required to treat visual images as output for a promising PMPR model. Recently developed deep-learning-based GAN techniques such as a deep convolutional GAN (DCGAN), a cycle-consistent GAN (Cycle GAN), and a conditional GAN-based image to image translation (Pix2Pix) could be of great help via the creation of realistic microstructures. In this regard, we generated virtual micrographs for various types of steels using a DCGAN, a Cycle GAN, and a Pix2Pix and confirmed the generated micrographs are qualitatively indistinguishable from the ground truth.

Automated summary

The prediction of macro-scale materials properties from microstructures is crucial for research, and generative adversarial network (GAN) techniques can help in building high-quality PMPR models.