Machine learnings for CVD graphene analysis: From measurement to simulation of SEM images

Abstract In this study, we introduce an approach that applies machine learning (ML) in various procedures to predict graphene growth pattern in chemical vapor deposition (CVD) system. At first, CVD experiments were conducted to synthesize graphene using CH4 as a precursor on a Cu substrate at high temperatures, to get experimental data for training those models. Then, the size, coverage, domain density, and aspect ratio of graphene, which vary depending on the synthesis conditions, were measured and analyzed automatically by developing a region proposal convolutional neural network (R-CNN). Subsequently, an artificial neural network (ANN) and a support vector machine (SVM) were used to develop surrogate models to deduce the correlation between CVD process variables and the measured specifications. Characteristic graphene grains with hexagonal morphology were created using a generative adversarial network (GAN) to imitate the CVD growth. Then, they were modified to have the same size and aspect ratio as the predicted values and placed to meet the predicted coverage and domain density. Finally, the pre-generated images were modified using Pix2pix to obtain the same outlook as the experimental SEM images. As a result, it was possible to simulate graphene synthesis under various CVD condition. Through numerous simulations in advance, we were able to identify the experiment condition to synthesize graphene with the desired morphologies of large grain size and low domain density. Developing a platform to predict a CVD system for the controlled synthesis of graphene allow us to synthesize the graphene with high efficiency, saving tremendous amounts of time and expenses.