3D Deep Learning Enables Accurate Layer Mapping of 2D Materials

Abstract

Layered, two-dimensional (2D) materials are promising for next-generation photonics devices. Typically, the thickness of mechanically cleaved flakes and chemical vapor deposited thin films is distributed randomly over a large area, where accurate identification of atomic layer numbers is timeconsuming. Hyperspectral imaging microscopy yields spectral information that can be used to distinguish the spectral differences of varying thickness specimens. However, its spatial resolution is relatively low due to the spectral imaging nature. In this work, we present a 3D deep learning solution called DALM (deep-learning-enabled atomic layer mapping) to merge hyperspectral reflection images (high spectral resolution) and RGB images (high spatial resolution) for the identification and segmentation of MoS 2 flakes with mono-, bi-, tri-, and multilayer thicknesses. DALM is trained on a small set of labeled images, automatically predicts layer distributions and segments individual layers with high accuracy, and shows robustness to illumination and contrast variations. Further, we show its advantageous performance over the state-of-the-art model that is solely based on RGB microscope images. This AI-supported technique with high speed, spatial resolution, and accuracy allows for reliable computer-aided identification of atomically thin materials.

Hyperspectral Fingerprints for Atomic Layer Mapping of Two-Dimensional Materials with Single-Layer Accuracy

“…Although what features the machine learning networks have learned are not totally explainable, this technique has shown satisfying performances for 2D materials’ thickness identification, especially suitable for initial screening of the microscopic images. Both supervised and unsupervised neural networks have been adopted to resolve the classification, segmentation, − and clustering analysis tasks using the microscopic images of 2D materials. The advantage of the machine-learning-based method is the one-time effort based on large amounts of data sets, which means that once the network is well trained, other researchers need merely a small number of their own data to fine-tune the network parameters before usage.…”

Applied Machine Learning for Developing Next‐Generation Functional Materials

“…© 2021 Wiley-VCH GmbH has meant that many other areas of research have also benefited, especially those of modular/combinatorial nature such as metal organic frameworks (MOFs), [103][104][105][106][107] covalent organic frameworks (COFs), [103,108] and 2D materials, [109][110][111][112][113] to name a few. For example, Aspuru-Guzik, et al recently harnessed a variational autoencoder to design new MOFs for the CO 2 separation.…”

Deep‐Learning‐Based Microscopic Imagery Classification, Segmentation, and Detection for the Identification of 2D Semiconductors

“…[37][38][39] However, to the best of our knowledge, there is no comprehensive study on three crucial computer vision tasks (i.e., classification, segmentation, and detection) in 2D materials. Generally, a classification model can predict the presented categories of flakes (single-label or multilabel); [40] a semantic segmentation model can generate a pixelwise segmentation map for the existing categories; [41][42][43] a detection model can classify and localize different 2D flakes using bounding boxes around the objects of interests. [44] Although previous studies realized high accuracy after network training with collected data of various 2D materials, they mainly focused on a rough identification of thickness such as mono-, few-, and thick-categories without using common datasets for comprehensive studies.…”