GAN-Based Realistic 3D Multicrystalline Si Structure Generation Using Actual Si Ingot Crystallographic Information

We have developed a method to generate realistic 3D multicrystalline (mc) structures based on one of the image generation AI; Generative Adversarial Networks (GAN). This was made possible by our original data science approach that can analyze Si ingot at the ingot scale, collecting crystallographic features such as grain boundary distribution, grain shape, crystal orientation, growth direction, and height. Hereby, our GAN-based model can learn and generate a variety of mc-structures containing actual crystallographic features, which are extremely important for understanding the properties and behavior of mc-materials. For example, in Si, crystal orientation is extensively involved in twinning and stress distribution around grain boundaries. Therefore, it is important to elucidate the relationships between mc-structures and properties through extensive exploration for mc-structures that include crystallographic features. However, it has been limited to localized and time-consuming exploration due to the difficulty of mc-structure reproduction. For instance, property estimation is based on mc-structures represented by 3D Voronoi diagram and physical simulations rely heavily on researchers’ sense and experience. This is where our method comes into play. Our method can overcome the mc-structure reproduction difficulty and enable faster and more comprehensive mc-structures exploration and analysis.<br/>We used 50 wafers of 15.6 cm × 15.6 cm × 180 µm sliced from a high performance (HP) mc-Si ingot grown via unidirectional solidification method. Alkaline texturing was performed to leave mainly the {111} facets which form a pyramid-like structure. Light illumination on these pyramids results in the appearance of crystallographic orientation-specific reflection patterns. We captured these reflection patterns by a home-made apparatus. As the light rotates around the sample, optical images were captured and integrated into the signal intensity matrix, which we call the “luminance profile”. Ingot scale crystal orientation mapping was performed by inputting luminance profiles into our previously developed crystal orientation estimation model that consists of a long short-term memory (LSTM) neural network and two fully connected layers. Our data science approach made it possible to obtain sufficient amount of ingot scale crystallographic feature data for machine learning, which was difficult by conventional measurement such as electron backscatter diffraction (EBSD) due to its slow measuring speed and small area. We adopted one of the GAN models StyleGAN2-ADA to learn crystallographic features since it can be trained effectively with limited data. This is because it applies a method called Adaptive Discriminator Augmentation (ADA), which allows for adaptive data expansion depending on the training process stage. We trained the model using 2500 randomly cropped to 2.6 cm square crystal orientation images from 50 images of 15.6 cm squares. After training, a variety of crystal orientation images can be generated from arbitrary noises. Moreover, we selected one of the conditional GAN models pix2pix to learn growth direction and height features which are especially important for materials grown via unidirectional solidification method. We trained the model using 1500 crystal orientation image pairs equally spaced 180 µm apart along the growth direction. After training, the model learns the growth behavior along the growth direction and can predict the crystal orientation image 180 µm above the input. Finally, realistic 3D mc-Si structure can be generated by combining trained StyleGAN2-ADA and pix2pix. This involves generating the bottom most crystal orientation image from noise and then repeating 180 µm upper crystal orientation image prediction and stacking them. Further analysis of the certainty and accuracy of the mc-structures generated by this method will be reported.