Electron microscopy and machine learning converge: Advanced analysis method for microstructures developed by NIMS team

Molybdenum disulfide (MoS₂) is a new material consisting of a few atomic layers with excellent semiconductor properties, attracting worldwide attention as a material for next-generation electronic devices. The performance of this material depends on microstructures such as the presence or absence of nanometer-level minutely rotated (twisted) regions and polarity (direction of atomic arrangement). It has been difficult to evaluate these microstructures with high accuracy over a wide area using conventional technologies, leaving ideal material design and manufacturing process adjustments in a trial-and-error state. To accelerate innovative material development and device applications in the future, new analytical techniques capable of analyzing such twists and polarities at the nano level are indispensable.

A research team consisting of researchers of the Center for Basic Research on Materials at NIMS, including Director Koji Kimoto, Principal Researcher Koji Harano, Principal Researcher Jun Kikkawa, Senior Researcher Ovidiu Cretu, Special Researcher Yoshiki Sakuma, and Principal Engineer Fumihiko Uesugi, in collaboration with Professor Yoshifumi Oshima and Senior Lecturer Kohei Aso from the Japan Advanced Institute of Science and Technology (JAIST), and Group Leader Takashi Matsumoto at Tokyo Electron Technology Solutions Ltd., has developed a method capable of analyzing the twists and polarities of monolayer MoS₂ in nano regions by combining the latest 4D scanning transmission electron microscopy (4D-STEM) with machine learning.

First, monolayer MoS₂ thin films were synthesized on sapphire substrates using metal-organic chemical vapor deposition (MOCVD), which is also used in semiconductor processes. A large number (more than 20,000 points) of diffraction patterns was acquired with nanometer spatial resolution using 4D-STEM. The enormous diffraction pattern data (4.6 GB) obtained was analyzed using nonnegative matrix factorization and hierarchical clustering, which are types of unsupervised machine learning. The diffraction intensities were converted to electron counts to estimate quantum noise, and the polarity and diffraction patterns of monolayer MoS₂ films were confirmed through simulations and experiments. The team succeeded in visualizing the polarity of twist domains from the noncentrosymmetry of diffraction patterns due to the violation of Friedel's law.

The data measurement and analysis used DigitalMicrograph electron microscopy software, along with Python libraries running on it, and original scripts were developed. Some of the independently developed software was made available on the following website: https://www.nims.go.jp/AEMG/DMindex.html.

The collaboration between 4D-STEM and unsupervised machine learning enables simultaneous capture of positional information on specimens and atomic arrangement information from diffraction patterns, and the effects of specimen curvature and other factors that have made analysis difficult can now be analyzed through clustering. This method has made it possible to visualize crystal growth processes and overall defect distributions over wide areas, which was difficult with conventional electron microscopy techniques. Unsupervised machine learning makes it possible to objectively extract features from unknown materials as well.

The measurement method developed in this study can be applied to the evaluation of various composite materials as well as to two-dimensional materials and may accelerate materials science and device development. Furthermore, by combining the advancement of electron microscopy equipment and the optimization of data analysis algorithms, particularly the knowledge of advanced measurement techniques, with machine learning, more sophisticated material evaluation will become possible, and the development of wide-ranging applications is expected in both industry and academia.

Journal Information
Publication: Small Methods
Title: Unveiling Twist Domains in Monolayer MoS₂ through 4D-STEM and Unsupervised Machine Learning
DOI: 10.1002/smtd.202501065