Automated Graphene Characterization Using Unsupervised Machine Learning
Graphene-based heterostructures are known to display many novel and exotic quantum phenomena. A necessary step in fabricating these devices is the characterization of graphene which is done manually using an optical microscope. With the help of machine learning algorithms, this process can be automated. Thus, it will no longer require a person to be physically present in the laboratory, which is especially important during these pandemic times. Graphene image analysis involves extracting the green value of the pixels (G value) from the images of graphene on a silicon oxide substrate taken by an optical microscope. The extracted value is subsequently used for the characterization of monolayer and bilayer graphene flakes from the background and the residues on the substrate. The images are preprocessed by dividing them into subsections with a constant light intensity to avoid the detrimental effects of vignetting. In the process of analyzing the images, it was found that the pixels corresponding to graphene, substrate, and residues tend to form Gaussian-like clusters in the G value color space. Hence, the Gaussian Mixture Model with its Expectation-Maximization algorithm could be utilized to fit the multiple Gaussians from the data. The Gaussian mixture model was implemented in the MATLAB environment. We report that the program was able to fit multiple Gaussians that are present in the data and detect the peaks corresponding to graphene enabling us to correctly characterize graphene flakes.