Data augmentation-assisted transfer learning for efficient monolayer recognition

The rapid and precise identification of monolayers synthesized by mechanical exfoliation or Chemical Vapor Deposition (CVD) is essential for investigating two-dimensional (2D) materials. However, conventional characterization techniques typically rely on extensive manual analysis and large-scale annotated datasets, which constrain their efficiency in real-time monitoring and morphological analysis. In this work, we present a deep learning-assisted approach, termed the 2D Materials View Neural Network (2DMViewNet), designed for efficient monolayer recognition and crystal orientation analysis with a minimal number of training samples. Leveraging data augmentation-assisted transfer learning, 2DMViewNet reliably identifies monolayer samples using only 10 training images. Furthermore, an automated algorithm is proposed to analyze the orientations of triangular monolayers by extracting deep image features, providing critical insights into growth dynamics. To further explore the fundamental growth mechanisms, 2DMViewNet is integrated with an in-situ optical observation micro-CVD system for real-time investigation of the crystal growth process. This work not only provides a rapid and data-efficient approach for monolayer identification but also presents a powerful tool for probing the fundamental mechanisms underlying 2D material growth, ultimately facilitating the development of next-generation nanomaterials.