Machine-learning approach for operating electron beam at KEK e−/e+ injector Linac

Aug 3, 2026

Upload by Hazel

Modern accelerators require continuous adjustment of many parameters and monitored beam observables to reach stable target performance, yet expert manual tuning is time-consuming and hard to reproduce, especially for single-pass linacs. This study investigates machine-learning based accelerator tuning at the KEK e−/e+ injector Linac using Bayesian optimization, tree-structured Parzen estimator, and CMA-ES. The optimization targets maximizing electron-beam charge while reducing the energy-dispersion function, achieving performance comparable to skilled expert tuning while addressing environmental drift sensitivity.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages12

Reading Time30 min

AccessFree to download

Introduction

Machine-learning based accelerator tuning motivation

Bayesian optimization and unsupervised tuning goal

Evaluation algorithms and beam-tuning targets

arXiv :2401 . 14739v1 [physics .acc-ph] 26 Jan 2024 Machine-learning approach for operating electron beam at KEK e − /e+ injector Linac Gaku Mitsuka, 1, 2, ∗ Shinnosuke Kato,3 Naoko Iida, 1, 2 Takuya Natsui, 1, 2 and Masanori Satoh 1, 2 1 KEK, Oho, Tsukuba, Ibaraki 305-0801, Japan 2 SOKENDAI, Shonan Village, Hayama, Kanagawa 240-0193, Japan 3 The University of Tokyo, Bunkyo, Tokyo 113-0033, Japan (Dated: January 29, 2024) In current accelerators, numerous parameters and monitored values are to be adjusted and evaluated, respectively. In addition, fine adjustments are required to achieve the target performance. Therefore, the conventional accelerator-operation method, in which experts manually adjust the parameters, is reaching its limits. We are currently investigating the use of machine learning for accelerator tuning as an alternative to expert-based tuning. In recent years, machine-learning algorithms have progressed significantly in terms of speed, sensitivity, and application range. In addition, various libraries are available from different vendors and are relatively easy to use. Herein, we report the results of electron-beam tuning experiments using Bayesian optimization, a tree-structured Parzen estimator, and a covariance matrix-adaptation evolution strategy. Beam-tuning experiments are performed at the KEK e − /e+ injector Linac to maximize the electron-beam charge and reduce the energy-dispersion function. In each case, the performance achieved is comparable to that of askilled expert. I. INTRODUCTION To improve or maintain the high performance of modern accelerators, dozens or even hundreds of parameters must be optimized to accommodate the volatile conditions. Values monitored to determine the success or failure of the optimization include those of the beam orbit, beam charge, energy-dispersion function, emittance, and charge loss in each accelerator sector. Hundreds of values are monitored. Determining the operating parameters, such as the magnetic field, based solely on the beam dynamics is typically challenging. For example, the KEKe − /e+ injector Linac (referred to as the KEK Linac) has no monitors to diagnose the beam energy in each sector, and the energy gains of individual RF cavities are accurately determined only occasionally. In addition, changes in the environmental temperature affect the RF system and set energy drifts. Therefore, the actual operation requires beam-parameter optimization while the beam conditions are monitored. Hitherto, operation experts have optimized the beam parameters based on their knowledge and experience. Complex and sensitive accelerator operations, such live optimizations based on expert inputs, may be time consuming reproduce, even if the results satisfy the required criteria. In particular, in the case of Linac accelerating a beam in a single pass, a self-feedback mechanism does not exist, unlike the ring where the beam orbits. Thus, the beam condition cannot be reproduced easily even if the same operating parameters are set. Accelerator tuning using machine learning has recently garnered attention as an alternative to expert-dependent optimization. Machine learning has progressed significantly in terms of speed, sensitivity, and application range since the 2010s. Various libraries are available from ∗ [gaku.mitsuka@kek.jp](gaku.mitsuka@kek.jp) different vendors and are relatively easy to use. For example, a neural network approach learns the response between the applied current of a coil (input) and a beam orbit (output) over weeks to months. Subsequently, it predicts the best applied current to achieve an optimal beam orbit. Because neural networks are based on supervised learning, their regression calculation accuracy is generally higher than that of unsupervised learning. Additionally, they can detect (classify) anomalies. However, the overtraining problem, in which the regression accuracy decreases significantly when the test situation differs fr

FAQs

为什么传统的专家手动调参在加速器运行中会遇到限制？

现代加速器需要优化大量参数并满足多项监测指标，且环境温度等条件会引起漂移；单次试验成本较高、结果也难以复现，导致手动现场优化逐渐触及效率上限。

这项研究在KEK e−/e+ injector Linac上主要优化哪些性能指标？

通过多次参数变化进行调优，以最大化电子束电荷，并同时降低能量色散函数，从而提升束流传输与能量质量。

研究评估了哪些机器学习/优化算法用于加速器调参？

文中评估了贝叶斯优化（Bayesian optimization）、树结构的Parzen估计器（TPE）以及协方差矩阵自适应进化策略（CMA-ES）。

Aug 3, 2026

Upload by Hazel