Two‐Stage GAN‐Based Generation of Virtual 3D Multicrystalline Silicon Reproducing Nucleation and Crystal Growth Processes Using Crystallographic Information from Real Ingots

We have developed a method to generate a variety of virtual three‐dimensional multicrystalline structures based on one of the image generation AI known as generative adversarial networks (GANs). This method consists of two GAN models, StyleGAN2 with adaptive discriminator augmentation and pix2pix. The training was made possible by our original data science approach that can analyze silicon ingot at an ingot scale, collecting crystallographic features such as grain boundary distribution, grain shape, and crystal orientation. Our GAN‐based model can learn and generate a variety of multicrystalline structures containing actual crystallographic features, which are extremely important for understanding the properties and behavior of multicrystalline materials.