Automatic Identification of 2D Materials Based on Machine Learning Method

Wang, Quanjin

Undergraduate Honors Thesis

Automatic Identification of 2D Materials Based on Machine Learning Method Public Deposited

Analytics

Download PDF

Citations

Citeable URL: https://scholar.colorado.edu/concern/undergraduate_honors_theses/4b29b744h

Abstract

As highly anticipated quantum materials of low dimensionality, 2D materials have recently seen significant advances in their fabrication, characterization, modification, and application in research. The research on 2D materials is fascinating, and one of the most attractive directions relates to the creation of heterostructures with on-demand and unique quantum properties through the stacking of various 2D materials. However, finding and identification of 2D materials are challenging as the essential part of 2D material research. Although we can artificially estimate the number of 2D materials layers based on optical microscopy (OM), looking for sufficient and suitable 2D materials for further research is still time-consuming and labor-intensive. Given the excellent performance of machine learning in the field of computer vision, we design a method to rapidly and accurately identify 2D materials upon OM. We utilize the architecture of the Mask Region-Based Convolutional Neural Network (Mask R-CNN) to complete the model training with datasets of 2D materials images. Additionally, we optimize hyperparameters to maximize the evaluation results of the trained models. Subsequently, we successfully apply the trained models to predict optical images of 2D materials, outputting the prediction image with bounding boxes (position), segmentations (shape), and categories (monolayer, few-layers, and thick-layers). Based on the feasibility of models and algorithms, we program software for 2D materials identification for universal and convenient usage. Finally, we design a Machine-Learning-Based 2D Material Transfer System with a system software, that successfully achieves real-time identification of 2D materials, processes large batches of images, and realizes precise control of the motorized stage through simple usage of the system software. This thesis is upgradeable; we also discuss potential development directions in datasets, the function of 2D materials auto scanning, and different architectures of machine learning.

Creator