Deep Learning Assisted Optical Identification of Exfoliated Two-Dimensional Crystals

Up to now, hundreds of two-dimensional materials are being studied in the fields of condensed matter physics, material sciences and electrical engineering. The overwhelming approach to obtain 2D crystals in laboratory is a combination of the mechanical exfoliation and the exhaustive search under an optical microscope by a well-trained researcher. Here we report a generic flake-hunting approach assisted by deep learning that can achieve the automatic, real-time, accurate, and robust optical identification of the type and the thickness of various 2D crystals. A semantic segmentation method using the encoder-decoder convolutional neural networks (SegNet) was developed and trained to identify the type and the thickness of the mechanically exfoliated 2D crystals on a SiO₂/Si wafer. Besides the commonly used parameters such as the optical contrasts of the 2D crystals, deep graphical features can also be extracted and harnessed by the SegNet for accurate and robust identification. Our proposed method can be used for a wide range of research topics where initial screening and identification of nanomaterials are necessary.