Automating Semiconductor Research with Machine Learning
Developing new thin semiconductor materials requires quantitative analysis of a large amount of reflection high-energy electron diffraction (RHEED) data, which is time-consuming and requires expertise. To address this issue, scientists at Tokyo University of Science are identifying machine learning techniques that can help automate the analysis of RHEED data. Their findings could dramatically speed up semiconductor research and pave the way for faster, energy-efficient electronic devices.
The semiconductor industry has grown steadily since its first steps in the mid-twentieth century, and with the high-speed information and communication technologies it enabled, it has given way to the rapid digitization of society. Today, alongside tight global energy demand, there is a growing need for faster, more integrated, and more energy-efficient semiconductor devices.
However, modern semiconductor processes have already reached the nanometer scale, and the design of new high-performance materials now involves the structural analysis of semiconductor nanofilms. Reflection high-energy electron diffraction (RHEED) is a widely used analytical method for this purpose. RHEED can be used to determine the structures that form on the surface of thin films at the atomic level and can even capture structural changes in real time as the thin film is synthesized!
Make machine learning do most of the work
Unfortunately, despite all its advantages, RHEED is sometimes hampered by the fact that its output patterns are complex and difficult to interpret. In virtually all cases, a highly skilled experimenter is needed to make sense of the huge amounts of data that RHEED can produce in the form of diffraction patterns. But what if we could have machine learning do most of the work when processing RHEED data?
A team of researchers led by Dr. Naoka Nagamura, visiting associate professor at Tokyo University of Science (TUS) and senior researcher at the National Institute of Materials Science (NIMS), Japan, worked on just that. . In their latest study, published online June 09, 2022 in the international journal Science and Technology of Advanced Materials: Methods, the team explored the possibility of using machine learning to automatically analyze RHEED data. This work, which was supported by JST-PRESTO and JST-CREST, was the result of joint research by TUS and NIMS, Japan. It was co-authored by Ms. Asako Yoshinari, Prof. Masato Kotsugi also from TUS and Dr. Yuma Iwasaki from NIMS.
Surface superstructures are at the heart of materials science.
The researchers focused on the surface superstructures that form on the first atomic layers of clean single-crystal silicon (one of the most versatile semiconductor materials). depending on the amount of adsorbed indium atoms and slight temperature differences. Surface superstructures are atomic arrangements unique to crystal surfaces where atoms stabilize in different periodic patterns than those inside the bulk of the crystal, depending on differences in the surrounding environment. Because they often exhibit unique physical properties, surface superstructures are of great interest in materials science.
First, the team used different hierarchical clustering methods, which aim to divide the samples into different groups based on various measures of similarity. This approach is used to detect how many different surface superstructures are present. After trying different techniques, the researchers found that Ward’s method could better track actual phase transitions in surface superstructures.
Accelerating R&D of next-generation semiconductors
Scientists then sought to determine the optimal process conditions for synthesizing each of the identified surface superstructures. They focused on the indium deposition time during which each superstructure was most extensively formed. Principal component analysis and other typical methods of dimensionality reduction have not yielded good results. Fortunately, non-negative matrix factorization, a different clustering and dimensionality reduction technique, could accurately and automatically obtain the optimal deposition times for each superstructure.
Excited by these results, Dr. Nagamura remarks, “Our efforts will help automate work that typically requires time-consuming manual analysis by specialists. We believe our study has the potential to change the way materials research is done and allow scientists to spend more time in creative pursuits.
Taken together, the results reported in this study will hopefully lead to new and effective ways to use machine learning technique for materials science, a central topic in the field of materials informatics. This, in turn, would have implications in our daily lives as existing devices and technologies are upgraded with better materials. “Our approach can be used to analyze superstructures grown not only on single-crystal surfaces of thin-film silicon, but also on metallic crystal surfaces, sapphire, silicon carbide, gallium nitride, and various other important substrates. Thus, we expect our work to accelerate the research and development of next-generation semiconductors and high-speed communication devices,” concludes Dr. Nagamura.