As particle accelerators become increasingly complex, machine learning (ML) emerges as a solution to address various operational challenges. This article reviews recent applications, emphasizing that ML should carefully complement conventional physics-based models and covers three key areas: 1) performance optimization using Bayesian optimization and reinforcement learning; 2) high-efficiency beam diagnostics, such as Bayesian algorithm execution for emittance measurement and neural-network-based virtual diagnostics of phase-space distribution; and 3) anomaly detection using Bayesian inference or PCA (Principal Component Analysis)/SVM (Support Vector Machine) for predictive maintenance. These data-driven methods are becoming indispensable for enhancing accelerator performance and reliability.

Advances in deep learning over the past decade have evolved neural networks, which were simple pattern recognition models, into gigantic language models capable of solving complex intellectual tasks. In the first half of this manuscript, we will explain a deep learning model called a transformer, which is essential for realizing large-scale machine learning. In the second half, we will discuss the potential applications of machine learning to the natural sciences, citing several examples. We will consider from several perspectives whether deep learning, which is said to be data-driven and highly black-box modeling, can make a fundamental contribution to science.

The KEK e^－/e⁺ injector linac (LINAC) supplies electron or positron beams to accelerators in KEK (electron beams to SuperKEKB HER, PF and PF-AR, and positron beams to SuperKEKB LER). During accelerator operation, various parameters are continuously optimized to ensure stable operation and maximize performance, such as injection efficiency, beam emittance, dispersion, beam background, etc. To improve operation tuning performance, we have developed an accelerator tuning scheme by applying Machine Learning. Our R&D studies are based on accumulated LINAC operation data (control parameters, monitoring data, environmental data). This paper introduces and reports the current status of these studies.

Recent advances in machine learning technologies have been remarkable, with neural networks in particular evolving rapidly and becoming indispensable in our daily lives. In the field of accelerator physics, the integration of machine learning has also been actively explored in recent years. This article introduces the author’s research on applications of neural networks in ion sources and Low Energy Beam Transport (LEBT) systems. The first example highlights the application of dimensionality reduction techniques to an ECR ion source and LEBT. By reducing the number of parameters to be tuned using dimensionality reduction, tuning efficiency was improved in the LEBT system. However, in the case of the ECR ion source, performance actually deteriorated, suggesting that such techniques may be difficult to apply in systems where small changes can significantly affect beam conditions. The second example focuses on predicting beam intensity from the ECR ion source using image analysis. By using plasma emission images as input to a neural network, prediction accuracy of beam intensity was improved. It was also found that the model could adapt to changes in system conditions over long-term operation through retraining. Overall, neural networks show promise as tools for uncovering parameter correlations and serving as diagnostic instruments, and further development in this area is anticipated.

We are developing an automated accelerator tuning framework based on machine learning (ML) at the XFEL facility, SACLA. SASE-XFELs are inherently unstable from shot to shot, making manual tuning challenging and highly dependent on operator expertise. This conflicts with the growing demand for diverse, advanced XFEL properties. To address this, we have developed and implemented an automated tuning system based on Gaussian Process (GP) Bayesian optimization (BO). This optimizer is now fully utilized in daily operations. Beyond simple pulse energy maximization, we have expanded its application to other XFEL properties such as spectral brightness, spectral shaping, and spatial profile. It is also applied to reproduce the beam condition at the injector section in the upstream. On-going developments are focused on advancing both the performance metrics and the ML algorithms. An X-band transverse deflector system is under development to provide longitudinal bunch information, enabling the optimization of XFEL pulse duration. The modular architecture of this framework is also being upgraded to incorporate other sophisticated algorithms.

Accelerator facilities require the control of numerous parameters; at the REKEN RI Beam Factory (RIBF), a complex of cyclotrons and linacs, more than 600 parameters—including environmental factors—affect beam quality. To optimize them more efficiently, we have been developing Bayesian optimization (BO) techniques, focusing on indices suitable for high-intensity beams and methods that maintain operational safety. We established a technique to measure beam transmission and spot size simultaneously using charge-converted particles downstream of the target, and we are also studying line BO with safety functions, so called SafeLineBO. This article discusses the major challenges, practical considerations, and failure cases encountered when applying machine-learning-based optimization to real accelerator operations.

By applying image recognition techniques based on Convolutional Neural Networks (CNN), we extracted numerical parameters such as injection momentum and injection timing offset, as well as distribution in phase space, from images known as mountain plots. In addition, using image generation techniques based on Conditional Variational AutoEncoder (CVAE), we succeeded in generating mountain plot images that cannot be obtained through direct measurement. This report describes how these image recognition and generation techniques were utilized to achieve these results.

This article describes research on hypernuclei using nuclear emulsion, a tracking detector for charged particles. By incorporating deep learning generalization capabilities into analysis, the exploration and automatic detection of hypernuclear events have been enabled. This revival of a historical and mature detector is advancing research into baryon-baryon interactions.

The maximum accelerating gradient of superconducting RF (SRF) cavities is limited by defects, making optical inspection of the inner surface an indispensable process. This research developed high-speed defect detection software to support manual inspection by reducing the number of images to be checked. For defect detection, a two-stage object detection model Faster R-CNN was employed. Training utilized a pre-trained model, enhancing detection performance through data augmentation (rotation and inversion) and fine-tuning. Inference was executed via parallel processing using GPGPU, yielding detection results for an entire 9-cell cavity in approximately 3 seconds. Performance evaluation of this software confirmed that its recall rate exceeded that of novice human inspector by more than 30%, it could detect a defect in a brand-new cavity completed after training data creation, and it enabled real-time inspection at around 10 frames per second. These results confirm that the developed system possesses performance comparable to existing manual inspection while significantly reducing working time. This research is considered to have high academic and engineering value as an example demonstrating the integrated application of image processing and machine learning in quality assurance for SRF accelerating cavities.

In recent years, there have been numerous examples of successful applications of machine learning to accelerator control and operation analysis. This article aims to provide broadly useful information, albeit not in great detail, on accelerator equipment control and the handling of various types of big data, targeting a wide range of people, from those just starting to introduce machine learning into accelerator operations to those who have already made significant progress in its implementation.

IPAC rotates among Asia, Europe, and the Americas, and in 2025, it was held in Taipei, Taiwan. This report summarizes both personal experiences and authors’ selections of key developments at the conference. Highlights include discussions on sustainability of future accelerators, developments of photocathodes, and importance of connecting with others without borders.

For the sustainable growth of the accelerator science field, educational programs for the next generation are essential. To this end, we launched a program called “AxeLatoon.” The program started in 2020, and activities such as constructing small-scale cyclotron accelerators are being conducted at several KOSENs (National Institutes of Technology). In the Special Lecture and poster session of the 22nd Annual Meeting of PASJ, KOSEN students involved in this program presented their activities. This paper reports on those efforts.

The Accelerator and Laser Young Researchers School was held on September 14–15, 2025, in Hiroshima with 47 participants, including students, young researchers, and invited lecturers. The program provided introductory lectures on accelerator physics and laser science, aimed especially at first-year graduate students, and offered opportunities for networking across different fields. Feedback confirmed that the school was a valuable occasion for learning fundamentals and building connections among the next generation of researchers. This report summarizes the program and presents voices from the participants.