AI Analysis Method Developed to Visualize Electronic States from Short-Duration Measurement Data Without Prior Training

Key Points of the Announcement

- In advanced scientific measurements such as those using synchrotron radiation, acquiring large amounts of training data has been difficult, posing a challenge for the introduction of AI analysis.

- This research has developed an AI analysis method that does not require prior preparation of training data, demonstrating its ability to clearly visualize the electronic band structure from a single, short-duration soft X-ray ARPES image.

- This achievement is expected to serve as a new AI analysis foundational technology that can overcome the challenge of not being able to prepare ground truth data in advance for advanced scientific measurements targeting unknown phenomena.

【Overview】

A research group comprising Yuichi Yokoyama and Kohei Yamagami from the Japan Synchrotron Radiation Research Institute (JASRI), Yuta Sumiya (then a Master's student) and Professor Ikuo Ono from the University of Electro-Communications, and Professor Kenichiro Mizumaki from Kumamoto University has developed an artificial intelligence (AI) based analysis method that estimates signal components from short-duration scientific measurement data without using pre-trained data.

This method utilizes the property of deep neural networks known as Deep Prior. The network's learning process first reproduces the intrinsic signal pattern of the sample and then proceeds to reproduce periodic noise (artifacts) originating from the measurement apparatus and random noise. By stopping the learning at an appropriate timing, unnecessary noise and artifact components can be suppressed, and signal components can be extracted.

The research group applied this AI analysis method to soft X-ray Angle-Resolved Photoemission Spectroscopy (ARPES) at the SPring-8 BL25SU large synchrotron radiation facility, demonstrating that band structures can be clearly visualized even from short-duration measurement data containing random noise and grid-like artifacts. Furthermore, they established it as an AI analysis method with objectivity and reproducibility by introducing a mechanism to automatically determine the timing for stopping the learning.

This achievement is expected to be developed as a new analysis foundational technology for introducing AI for Science into experimental sites in advanced scientific measurements using synchrotron radiation, neutrons, etc.

This research result will be published in the international scientific journal 'Machine Learning: Science and Technology' published by the Institute of Physics on June 12th.

【Background of the Research】

In recent years, deep learning has achieved significant results in natural language processing and image analysis. However, many state-of-the-art AI models assume the availability of large amounts of training data, making them difficult to directly apply to advanced scientific measurements aimed at elucidating unknown phenomena where the ground truth is not known in advance. In particular, experiments using large research facilities such as those for synchrotron radiation and neutrons are costly per measurement, making it difficult to prepare a sufficiently large dataset for AI training.

Angle-Resolved Photoemission Spectroscopy (ARPES) is a powerful technique for directly probing the band structure, which is the relationship between the energy and momentum of electrons in a material. ARPES using soft X-rays from synchrotron radiation is particularly sensitive to the electronic states within a material and is useful for investigating the band structure in three-dimensional momentum space. On the other hand, it is prone to low signal-to-noise ratios (S/N), making it difficult to obtain high-quality band structure images in a short time. Additionally, in the Fixed mode, which measures efficiently by fixing the photoelectron analyzer voltage, grid-like artifacts unrelated to the intrinsic electronic states of the sample appear in the measurement images.

Suppressing these grid-like artifacts in addition to random noise, while visualizing the band structure originating from the intrinsic electronic states of the sample, has been a significant challenge in ARPES.

【Research Content and Results】

Development of AI Analysis Method

To overcome this challenge, the research group developed an AI analysis method that does not require prior preparation of training data. In this method, a small-scale deep neural network is optimized for each measurement image to be analyzed (Figure 1). Deep neural networks have a property that makes them more likely to learn characteristic structures and repeating patterns in images than random noise. This property acts as a prior distribution that preferentially reproduces the intrinsic structure contained in the image, and is called Deep Prior.

By utilizing Deep Prior, random noise and grid-like artifacts contained in short-duration measurement images obtained in Fixed mode can be separated, and sample-derived signal components can be extracted. The key to this method is a mechanism that automatically stops the AI's learning at an appropriate timing. If learning progresses too far, the AI will faithfully reproduce unnecessary noise and artifacts. Therefore, by combining the change in mean squared error between the AI output image and the measurement image with the intensity of grid-like artifacts, an algorithm was constructed to automatically determine the optimal stopping timing before artifacts begin to be reproduced, while suppressing the influence of noise. This enables the simultaneous suppression of complex artifacts and extraction of signal components, which was difficult with previous Deep Prior-based methods.

Demonstration with Soft X-ray ARPES Experimental Data from SPring-8

In this study, the effectiveness of the developed AI analysis method was verified using soft X-ray ARPES experimental data of CeRu2Si2, known as a heavy fermion material, acquired at the SPring-8 BL25SU large synchrotron radiation facility. The target Fixed mode measurement data included signal components originating from the sample's electronic states, as well as random noise such as photon shot noise and grid-like artifacts caused by the measurement apparatus's metal mesh.

When this method was applied to short-duration measurement data of 40 seconds and 10 seconds, it succeeded in separating grid-like artifacts and random noise and extracting band structure information (Figure 2). Compared to standard data obtained in Swept mode (measurement time 2880 seconds), where the influence of artifacts is averaged by the measurement device, it was revealed that the main features of the band structure could be visualized more clearly with this method. Furthermore, quantitative evaluation was performed using momentum distribution spectra extracted from the ARPES images. When the AI analysis method was applied to 40-second measurement data, it became possible to separate peak structures that overlapped in the conventional Swept mode, demonstrating that band structure information could be extracted in more detail even under measurement conditions approximately 70 times faster. On the other hand, although spectral shape distortion due to residual noise was observed in the 10-second measurement data, it was shown that qualitative band structure understanding could be achieved with a speed-up of nearly 300 times (Figure 3).

【Future Prospects】

This achievement will enable the visualization of band structures with short measurement times in soft X-ray ARPES at SPring-8, leading to improved experimental efficiency through reduced measurement time and decreased sample damage from synchrotron X-rays. Additionally, the developed AI analysis method can perform analysis in approximately 20 seconds. Therefore, if the measurement time is 20 seconds or longer, real-time AI analysis can be performed concurrently with a series of measurements, directly leading to improved measurement efficiency compared to conventional measurement times (several minutes to tens of minutes). In advanced materials research, such as for functional materials and quantum materials, it is important to investigate the electronic states within materials with high precision and efficiency.