Characterizing the heterogeneity of 2D materials with transmission electron microscopy and machine learning

Lee, Chia-Hao

Permalink

https://hdl.handle.net/2142/120260 Copy

Description

Title

Characterizing the heterogeneity of 2D materials with transmission electron microscopy and machine learning

Author(s)

Lee, Chia-Hao

Issue Date

2023-04-18

Director of Research (if dissertation) or Advisor (if thesis)

Huang, Pinshane

Doctoral Committee Chair(s)

Huang, Pinshane

Committee Member(s)

• Madhavan, Vidya
• Schleife, André
• Zuo, Jian-Min

Department of Study

Materials Science & Engineerng

Discipline

Materials Science & Engr

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• 2D materials
• scanning transmission electron microscopy
• machine learning
• heterogeneity

Abstract

Two-dimensional (2D) materials have brought countless wonders to the field of condensed matter physics, materials science, and nanotechnology for almost two decades since its inception. These materials exhibit a range of unique properties that make them ideal candidates for various applications, including flexible electronics, energy conversion, and catalysis. However, the properties of 2D materials can vary significantly due to their heterogeneity, which arises from the presence of defects, grain boundaries, interlayer twist angles, and other structural imperfections. Therefore, understanding the structure, distribution, and impact to different properties of such heterogeneity has become a major challenge in the field of 2D materials because such characterization would often require both high spatial resolution, measurement precision, and detection limit. While scanning and transmission electron microscopy (S/TEM) has been used to characterize the structural and chemical properties of 2D materials, a common issue associated with S/TEM techniques is the inherent trade-off between resolution and field-of-view. Additionally, radiation damage from the high energy electrons sets a practical limit to accessible 2D materials system, and also the achievable measurement precision. This thesis focuses on how to combine S/TEM techniques with machine learning algorithm to overcome the abovementioned trade-offs and challenges while characterizing 2D materials and their heterogeneities. The thesis begins with in situ study of phase transitions in a phase-change 2D material, MoTe2. We investigate the reversible phase transitions of MoTe2 from micro- to atomic scales using in situ TEM with graphene encapsulation. We use laser irradiation to convert few-layer MoTe2 flakes from the 2H to a mixture of 1T' and Td phases, and then use in situ pulsed heating to monitor the reverse phase transition. We observe a highly anisotropic phase transition from the Td to 2H phase and characterize the atomic structure of the phase boundaries, which play an important role in the phase transition kinetics. Next, we develop a machine learning framework to locate and classify each point defect in large data sets of atomic-resolution images of monolayer WSe2. This enables us to generate class-averaged images of single-atom defects with 0.2 pm precision, and allows us to experimentally visualize the pm-scale oscillating strain field induced by a single atom defects for the first time. We also extend the generalizability of this framework by generating high quality training data for machine learning applications using a cycle generative adversarial network (CycleGAN). By utilizing CycleGANs, which are commonly used for image-to-image translation tasks, we are able to transfer experimental imperfections from real data to simulated data, generating a high quality training set for defect identification tasks using fully convolutional network (FCN). Our results show that FCNs trained on this processed data achieve comparable defect identification performance to manually optimized training sets while requiring much less human intervention. Lastly, we achieve unprecedented levels of detail for structural disorder in 2D materials using deep sub-angstrom resolution electron ptychography. With a spatial resolution of 0.41 angstrom, more than 2x better than the conventional limit, we obtain sub-picometer precision measurements of defect position, strain distribution, and 3D local rippling of a monolayer WSe2 from a single ptychographic reconstructed image. Our approach opens up new opportunities for sub-angstrom resolution imaging of structural disorder, including interfaces and grain boundaries. In summary, this thesis demonstrates a combination of new S/TEM techniques with machine learning algorithm which enable atom-by-atom characterization of heterogeneity of 2D materials including phase boundaries, strain, point defects, and local rippling.

Graduation Semester

2023-05

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/120260

Copyright and License Information

Copyright 2023 Chia-Hao Lee

Owning Collections