Using Convolutional Neural Networks to Segment Scanning Electron Microscopy Images of Graphene

Scanning electron microscopy (SEM) is one of the most commonly used techniques to characterise graphene synthesised by chemical vapour deposition. These images contain detailed information about the crystalline quality, nucleation density, domain size, and percent coverage, but typically are analysed through laborious processes that require the trained eye of synthesis experts. To facilitate the use of data-driven techniques to optimise the synthesis recipes of graphene, we require analysis results from many images, which is very time and labour intensive. To overcome this problem, we use convolutional neural networks that, once trained, are capable of rapidly analysing such images. We present two such networks that can segment images of graphene and identify regions of the image that represent graphene. First, we utilise a U-Net architecture coupled with moderate image augmentation. This encoder/decoder architecture has been effective in segmenting biomedical images and can be trained on relatively small-sized datasets. It first down-samples the input image into a latent space representation with successive up-sampling. Each up-sampling layer is combined with input from the down-sampling steps to impart contextual information to the output. To adapt this for graphene, we added augmented 73 images with modified brightness, zoom, and shear. Using this approach we demonstrate an accuracy of 0.92, precision of 0.88, and recall of 0.95. Second, we demonstrate the promise of a neural network trained on images with domains of graphene represented as star-convex polygons. We use a U-Net architecture as the basis of our model, with an additional convolutional layer at the end. The information obtained from these models can then be used to provide accurate quantitative metrics such as percent coverage, nucleation density, and average domain size.