Apoorva Bisht

Author: Apoorva Bisht | Majors: Physics and Computer Science | Semester: Fall 2022

I joined Dr. Luu’s lab in November 2021 and started working on this project in January 2022. To begin with, I studied the basics of the detection model that was to be used for the project after which, I worked on getting an understanding of the data collection process. I am grateful to Dr. Luu and Xuan Bac Nguyen for their constant support and guidance through this project. The work done in spring and fall 2022 is as follows.

The goal of this ongoing project is to develop an application that can be utilized by different groups for varied materials to detect 2D material flakes and characterize their thickness hence, allowing researchers to feed multiple optical microscope images (or a live video) of flakes and the desired thickness range, and receive output indicating the potentially useful flakes. This eliminates the need for using an additional equipment like AFM to determine the thickness.

There were two main tasks for this project: (A) prepare dataset for specific materials and (B) use the dataset to train a model and further optimize it. Task A further involved three steps: (a) collect calibrated microscope images, (b) determine thickness of each flake using AFM and (c) annotate the flakes using CVAT (Computer Vision Annotation Tool). These tasks are explained as follows.

1. Since the AI model uses contrast and color of the flakes to distinguish the flakes from the substrate and to determine their thickness, it is important that flakes of specific thickness are captured as having similar RGB values independent of when the images are taken. Hence, we use calibration slides to first calibrate the microscope settings. After this, microscope images at 20x magnification are captured.
2. The next step involves gathering thickness information for the flakes captured in step (a). This is done via AFM (Atomic Force Microscopy). AFM allows us to measure thickness with nm precision. Determining thickness of a flake can take anywhere from 15-30 minutes depending on the resolution.
3. The last step in data preparation is to combine the results from steps (a) and (b) and generate a dataset that can be used as input to the AI model. We use a software called CVAT (Computer Vision Annotation Tool) in which the images from step (a) are uploaded, bounding boxes are created and thickness values from step (b) are added as metadata to each flake.

Task B was mainly performed during summer, where the data from CVAT software was imported and a detection model based on MASK-RCNN was trained. This model is capable of detecting flakes on Si/SiO₂ substrate.

The progress during 2022 laid the foundation for the project that will be continued in spring 2023. We envision that our work could deliver a significant impact by developing an application that can be utilized by researchers to considerably speed-up the flake hunting process. Utilization of such low-cost applications that exploit the color-to-thickness correspondence of 2D materials can allow researchers to find flakes of desired thickness in minimum time with minimum labor.