Shinji Kawaguchi (Assistant Professor, Department of Physics)

Hayato HASEGAWA (2nd Year Master's Student, Department of Physics)

Masamitsu Hayashi, Associate Professor, Department of Physics

Eishin Nakatani (Professor, The University of Electro-Communications)

Kenji Tanabe, Associate Professor, Toyota Technological Institute

Takuya Sawa (Toyota Technological Institute, 2nd year of master course)

Keisuke Yamada (Assistant Professor, Gifu University)

Key points of the presentation

• We succeeded in estimating the difficult-to-measure properties of materials using image recognition technology based on artificial intelligence.
• This is the first time in the world that multiple magnetic properties of a magnet can be estimated ^{from a} single magnetic domain image ^{(Note 1)}.
• Extending this method to various magnetic properties may significantly shorten the R&D period for next-generation information devices.

Summary of Presentation

With the rapid progress of informatization of society, the development of energy-saving and highly efficient information devices has become a major issue. Characterization of target materials and devices occupies an important position in the development of next-generation devices based on new materials and physical principles. However, until now, such characterization has required special measurements and micromachining, and has required a great deal of labor. The group led by Assistant Professor Masashi Kawaguchi, 2nd-year master's student Jun Hasegawa, and Associate Professor Masamitsu Hayashi of the Graduate School of Science, The University of Tokyo, has joined Professor Eishin Nakatani of the Department of Information and Network Engineering, Graduate School of Information Science and Engineering, The University of Electro-Communications, Associate Professor Kenji Tanabe and 2nd-year master's student Takuya Sawa of the Faculty of Engineering, Toyota Technological Institute, and Associate Professor Keisuke Yamada of the Graduate School of Science and Engineering, Gifu University. In collaboration with a research group led by Assistant Professor Keisuke Yamada of the Department of Chemical and Biomolecular Engineering at Gifu University, the research group was the first in the world to successfully estimate the magnetic properties of a magnet from its image using artificial intelligence. In this research, the magnetic properties of a nano-multilayer magnet, which is considered a candidate material for next-generation information devices, were estimated from a ^single magnetic domain image ^{(Note 1)} without further micromachining or measurement by combining machine learning, which has made great progress in the field of artificial intelligence, and computer simulation, Multiple magnetic properties were estimated from a single magnetic domain image (Note 1) without further micromachining or measurement. This achievement is expected to greatly accelerate the research and development of materials for the realization of innovative next-generation information devices.

Contents of presentation

Background of the research and problems in previous studies
In the ^{field of} spintronics ^{(Note 2)}, which aims to realize next-generation information devices, research and development is conducted based on the evaluation of parameters that describe the magnetic properties of materials (hereinafter referred to as "magnetic parameters"). However, special measurement methods are often required for their evaluation, which has been a challenge in research and development. For example, the magnetic parameter called the JarosinskiMoriya interaction ^{(Note 3) (} hereinafter abbreviated as DMI) is a typical example of such a parameter, and while it is the most important parameter in next-generation magnetic memory development research, no easy means of evaluating it has been established. However, there is no easy way to evaluate them. Therefore, there has been a strong demand for a simple and quick method to evaluate these magnetic parameters. In this study, we attempted to solve this problem by using machine learning, which has been rapidly developing in recent years.

Machine learning has developed dramatically with the progress of computer science and is now becoming an indispensable technology in modern society. Its applications are numerous, including automated driving, translation, image recognition, and the development of Go/Shogi (Japanese chess) software. Especially in the field of image recognition, the development of convolutional neural networks ^{(Note 4)} (hereinafter abbreviated as CNN) has realized highly accurate recognition and can capture subtle features that the human eye cannot notice. In this study, we attempted to use this CNN technology to analyze images containing magnetic information and obtain magnetic parameters such as DMI.

In a magnet, there are neat rows of compartments called "magnetic domains," in which the N poles are oriented in the same direction within each region. The size and shape of the magnetic domains reflect the magnetic parameters of the material, and if we can capture the characteristics of the magnetic domain pattern by acquiring and analyzing images mapping the N-pole orientation, we can infer the magnetic parameters.

Research content and results
Figure 2 shows an overview of this research. First, we prepare teacher data ^{(Note 5)} for the artificial intelligence to learn and test data ^{(Note 6)} to evaluate the accuracy of the learned artificial intelligence. Since a large amount of data is required for the teacher data, it is difficult to obtain it through experiments. Therefore, in this study, we used micromagnetic simulation ^{(Note 7)} to create magnetic domain images to serve as teacher data (Figure 1), and magnetic domain images obtained from experiments were used for test data to investigate whether DMI values ( D values ) could be estimated.

Figure 1: Example of magnetic domain pattern
An example of a magnetic domain pattern calculated by simulation. The white area shows a magnetic domain with the N-pole of magnetization facing outward toward the paper surface, and the black area shows a magnetic domain in the opposite direction. The number written in the lower right corner of the image is the magnitude of DMI ( D-value ), and ^the unit is mJ/m2. To the human eye, the structure appears to be random, but to the eye of an artificial intelligence, differences in magnetic domain patterns due to differences in D-value can be captured.

Figure 2: Conceptual diagram of this research
The artificial intelligence learns by using magnetic domain images obtained through numerical calculations (computer simulations) and the magnetic parameters used in the simulations. The goal of this research is to determine how accurately the artificial intelligence can estimate the magnetic parameters by reading experimentally measured magnetic domain images.

First, to confirm the effectiveness of machine learning using magnetic domain images, micromagnetic simulations were used to generate not only teacher data but also test data to estimate D values. We prepared 100,000 magnetic domain images as teacher data and 10,000 magnetic domain images as test data, independently from the teacher data. The D values of the teacher data and the test data were randomly varied. We can see that the estimated values are proportional to the set value with a slope of 1. Furthermore, the distribution of estimated values when the set values are fixed at 0.2, 0.4, 0.6, 0.8, and 1.0 ^mJ/m2 (Figure 3(b)) shows that the estimated D values are distributed in the neighborhood of the set values. These results indicate that machine learning using magnetic domain images is effective.

Figure 3: (a) Relationship between the set value of D-value and the estimated value. (b) Distribution of estimated values when the set value of D-value is fixed. (c) Measured D-values (red) and results estimated by the artificial intelligence (white and black) for different Ta film thicknesses. Two types of artificial intelligence were prepared (corresponding to white and black in the estimated results). The difference between the two artificial intelligences is the difference in the teacher data used for training. Specifically, the values of the magnetic parameter Keff set during the generation of the teacher data are different. Despite the differences in the teacher data, we found that the two AI returned nearly identical estimates of the D values and roughly reproduced the trends in the measured values. This suggests not only that artificial intelligence can successfully estimate D values, but also that it may be able to estimate the two magnetic parameters D value and Keff simultaneously.

Using the trained artificial intelligence, the next step was to estimate D-values using magnetic domain images obtained from experiments for the test data. The device structure used in the experiment is a multilayer film of Si sub./Ta ( d) /Pt (2.6 nm)/Co (0.9 nm)/MgO (2 nm)/Ta (1 nm). The magnet we are focusing on is a 0.9 nmCo ultra-thin film, but since the DMI ofCo is known to vary with the thickness of the lower Ta layer, we experimented with different Ta thicknesses. Magnetic domain patterns were observed using a microscope equipped with a small magnetic sensor. Analysis of the acquired images revealed that the D values measured from the experiment and those estimated by the artificial intelligence were in agreement, as shown in Figure 3(c). It was also found that the accuracy of the estimated values did not change when the number of magnetic parameters that could be estimated was increased. These results demonstrate that multiple magnetic parameters that are difficult to evaluate can be obtained from magnetic domain images without micromachining or electrical measurements by using machine learning and image recognition.

Significance of this research and future prospects
This study demonstrates that image recognition using machine learning is extremely effective for the characterization of materials, which has been difficult until now. In the future, if it is possible to successfully train an artificial intelligence by machine learning based on the generation of teacher data of magnetic domain patterns corresponding to various combinations of magnetic parameters using micromagnetic simulation, it will be possible to create a database of this information and obtain images of magnetic domains of materials and structures to be investigated by simply inputting the magnetic domain images. If we can successfully train the artificial intelligence to generate the corresponding magnetic domain pattern teacher data, we can create a database of this information and obtain all magnetic parameters simply by inputting the magnetic domain image of the material or structure to be investigated. This will revolutionize the approach to research and development by simplifying the characterization of materials that used to require a great deal of labor.

Journals

Journal name	npj Computational Materials
Title of paper	Determination of the Dzyaloshinski-Moriya interaction using pattern recognition and machine learning
Author(s)	Masashi Kawaguchi, Kenji Tanabe, Keisuke Yamada, Takuya Sawa, Shun Hasegawa, Masamitsu Hayashi, and Yoshinobu Nakatani
DOI Number	10.1038/s41524-020-00485-2
Abstract URL	https://doi.org/10.1038/s41524-020-00485-2

Terminology

Note 1 Magnetic domain image

When magnified, a magnet is divided into regions with aligned N-pole orientations. This region is called a magnetic domain, and the image mapping the N-pole orientation is called a magnetic domain image. ↑up

Note 2 Spintronics

Spintronics is a field of research and development that aims to realize new functions and high-performance devices that do not exist in conventional electronics by utilizing the spin degree of freedom of electrons. ↑up

Note 3 JarosinskiMoriya interaction

This is an interaction that facilitates the orientation of neighboring magnetizations in a spatially twisted direction of 90 degrees. It is actively studied as an important parameter in research for the development of next-generation magnetic memory, and is named after two researchers, Dr. Igor Jarosinski and Dr. Toru Moriya of Japan, who made major contributions to the discovery of this interaction around 1960. ↑up

Note 4 Convolutional Neural Network

One of the machine learning models for image analysis, it is famous for dramatically increasing the accuracy of image recognition. It is characterized by a model that emphasizes the closeness of vertical and horizontal data, rather than treating image data as one-dimensional data aligned vertically and horizontally. In English, it is written as Convolutional Neural Network, so it is sometimes called CNN for short. ↑up

Note 5 Teacher data

In machine learning, this refers to data consisting of a set of examples and answers for training artificial intelligence. By learning a large amount of this teacher data, the prediction accuracy of the artificial intelligence can be improved. In this experiment, the magnetic domain image data is the example and the magnetic parameters are the answers. ↑up

Note 6 Test data

In machine learning, data similar to teacher data used to evaluate the estimation accuracy of artificial intelligence. This test data must be independent of the teacher data. The estimation accuracy is revealed by feeding magnetic domain image data to the artificial intelligence and comparing the resulting magnetic parameters estimated by the artificial intelligence with the actual magnetic parameters. ↑up

Note 7 Micromagnetic simulation

This is a type of numerical calculation method (computer simulation) to investigate the structure of magnets. This method divides a magnet into microscopic sizes (cells) and solves the equations of motion for each cell while taking into account the interaction between the cells to investigate the time evolution. ↑up