A novel method to determine the layer number of 2D TMD materials based on Optical Microscopy and Image Processing

TL;DR

This paper introduces a rapid, easy-to-apply optical microscopy and image processing method using machine learning to accurately determine the layer number of 2D TMD materials, outperforming traditional techniques in speed and simplicity.

Contribution

The study presents a novel, single-training machine learning approach that accurately predicts layer numbers of 2D TMD materials using optical images, reducing the need for multiple training samples.

Findings

High accuracy in layer number prediction for MoS2, WSe2, WS2.

Method is faster and easier than Raman, AFM, PL.

Requires only a single training sample per material/substrate.

Abstract

Two-dimensional (2D) transition metal dicakcoganite (TMD) materials have unique electronic and optical properties. The electronic band structures of the materials alter as a function of layer numbers, which results in modifications to the whole characteristic properties. Therefore, the determination of the layer number is crucial for optoelectronic applications. In this study, a fast and easily applicable method is proposed for determining the layer number of two-dimensional TMD materials by using an optical microscopy and computatial methods. The method uses image processing techniques on digital images taken with a Complementary Metal Oxide Semiconductor (CMOS) camera under a conventional reflecting microscope. The chromaticity differences and lightness difference in International Commission on Illumination (Commission internationale de ,CIE) Luv color space values between the layered areas on the flakes and substrates are used together to train the model to be identified the layer numbers of the materials and corrected via photoluminescence spectroscopy. The random forest clasifier algorithm is applied to predict layer numbers on unknown materials. This approach provides high accuracy under varying microscope configurations such as brightnesses, gain etc. and material-substrate combinations. From ML to bulk whole layers numbers of MoS2, WSe2 and WS2 flakes are determined with high accuracy by training only a single flake of each has various layer numbers. Compared with traditional methods such as Raman spectroscopy, atomic force microscopy (AFM) and photoluminescence (PL), our method is not only faster but also easier to apply. Moreover, unlike the prominent methods in the literature such as machine learning or clustering-based, it requires only single training for a material/substrate combination and further accelerates the processes.