Fast and accurate robotic optical detection of exfoliated graphene and hexagonal boron nitride by deep neural networks. (20th April 2021)

• Record Type: Journal Article
• Authors: Shin, Young Jae; Shin, Wheemyung; Taniguchi, Takashi; Watanabe, Kenji; Kim, Philip; Bae, Sung-Ho
• Is Part Of: 2D materials. Volume 8:Number 3(2021)
• Publication Date: 2021-04-20
• Subjects: graphene -- hBN -- deep neural network -- automatic optical detection; Graphene -- Periodicals; Materials science -- Periodicals; Nanostructured materials -- Periodicals; 620.115
• DOI: 10.1088/2053-1583/abd72c
• Languages: English
• ISSNs: 2053-1583

Abstract

Abstract: Van der Waals (vdW) heterostructure constructed from atomically thin layered materials provides quantum material platforms with emergent physical phenomena and novel device applications. While stacking atomically thin vdW layers in combination with automate machine vision identification and semi-automated stacking have been demonstrated, a combination of machine learning based automatic detection/identification assembly capability is necessary to further advance vdW heterostructure fabrication. Here, we developed a new automatic optical detection technique with a deep neural network (DNN) incorporated into a motorized microscope that automatically scans entire silicon wafers to detect and identify two-dimensional (2D) materials. We demonstrated the automated combination on an optical microscope (OM) with a DNN algorithm that enables identification and classification of graphene with different sizes, shapes and thicknesses. For this purpose, we trained a representative DNN for object detection with approximately 1000 OM images, resulting in high accuracy in detection and classification. We further verified the effectiveness of graphene trained DNN (GT-DNN) in practice, by confirming the yield of flakes depending on exfoliation method. In addition, we also showed the transferability of our method to other 2D materials. Our pre-trained GT-DNN was used to train a detection model for hexagonal boron nitride with only a few samples, resulting in further improvement of the accuracy of DNN-based detection. Our experimental method can effectively be extended to automatic detection and assembly of a plethora of 2D materials.