Bridging the gap between machine learning and particle accelerator physics with high-speed, differentiable simulations

Aug 3, 2026

Upload by Riley

Machine learning is increasingly used to advance particle accelerator operations and experiment analysis, but training state-of-the-art models is constrained by limited beam time, high simulation costs, and large-dimensional optimization. This work presents Cheetah, a PyTorch-based high-speed differentiable linear beam dynamics code designed to generate training data efficiently and enable gradient-based optimization. Cheetah reduces simulation times by orders of magnitude and supports accelerator tuning, system identification, and reinforcement learning environments. It demonstrates five applications, including beamline tuning, system identification, physics-informed Bayesian optimization priors, and modular neural network surrogate modeling for space charge effects.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages17

Reading Time43 min

AccessFree to download

Introduction

Challenges in data generation for ML in accelerators

Cheetah: PyTorch-based high-speed differentiable simulations

Applications and examples of Cheetah usage

Bridging the gap between machine learning and particle accelerator physics with high-speed, differentiable simulations Jan Kaiser, 1,* Chenran Xu,2,† Annika Eichler, 1,3 and Andrea Santamaria Garcia2 1Deutsches Elektronen-Synchrotron DESY, Hamburg, Germany 2Karlsruhe Institute of Technology (KIT), Karlsruhe, Germany 3Hamburg University of Technology, 21073 Hamburg, Germany (Received 12 January 2024; accepted 17 April 2024; published 28 May 2024) Machine learning has emerged as a powerful solution to the modern challenges in accelerator physics. However, the limited availability of beam time, the computational cost of simulations, and the high dimensionality of optimization problems pose significant challenges in generating the required data for training state-of-the-art machine learning models. In this work, we introduce Cheetah, a PyTorch-based highspeed differentiable linear beam dynamics code. Cheetah enables the fast collection of large datasets by reducing computation times by multiple orders of magnitude and facilitates efficient gradient-based optimization for accelerator tuning and system identification. This positions Cheetah as a user-friendly, readily extensible tool that integrates seamlessly with widely adopted machine learning tools. We showcase the utility of Cheetah through five examples, including reinforcement learning training, gradient-based beamline tuning, gradient-based system identification, physics-informed Bayesian optimization priors, and modular neural network surrogate modeling of space charge effects. The use of such a high-speed differentiable simulation code will simplify the development of machine learning-based methods for particle accelerators and fast-track their integration into everyday operations of accelerator facilities. DOI: 10.1103/PhysRevAccelBeams.27.054601 I. INTRODUCTION Future particle accelerator experiments will place everincreasing demands on the performance and capabilities of particle accelerator operations and experiment analysis. In order to meet these demands, the research community is increasingly turning to machine learning (ML) methods, which have already demonstrated their ability to push the envelope of what is possible in the field of accelerator science [1–4] . One of the remaining challenges holding back this line of research is the demand for large amounts of data (including environment interactions) from these methods. Reinforcement learning (RL), for example, has already successfully been used to train intelligent tuning algorithms and controllers that can outperform the currently deployed black-box optimization algorithms and handcrafted controllers [1,5–7] . However, RL methods require many interactions with their target task to train a well-performing policy. For example, 6 000 000 samples were needed in [1] to successfully train a *jan.kaiser@desy.de †[chenran.xu@kit.edu](chenran.xu@kit.edu) Published by the American Physical Society under the terms of the Creative Commons Attribution 4.0 International license. Further distribution of this work must maintain attribution to the author(s) and the published article’s title, journal citation, and DOI. policy on a transverse beam tuning task. Orders ofmagnitude larger number of samples are also common with other RL applications [8,9] . The general scarcity of beam time makes collecting experimental data for ML methods, such as RL, a significant bottleneck. Gathering at least partial datasets in simulation can alleviate this problem, but existing accelerator simulation codes have mostlybeen developed with a focus on the design phase of accelerators, where high fidelity and physical correctness are critical, and computing times range from minutes to several hours for one simulation. Consequently, data collection with existing simulation codes becomes impractical with the growing demand for large datasets. In this paper, we introduce Cheetah, a PyTorch-based high-speed differentiable linear b

FAQs

Why is bridging machine learning and accelerator physics challenging?

ML methods require large training datasets and many environment interactions, while beam time is scarce and conventional high-fidelity simulations are computationally expensive.

What is Cheetah and what problem does it address?

Cheetah is a PyTorch-based high-speed differentiable linear beam dynamics code that accelerates simulations by orders of magnitude and provides differentiable models for efficient gradient-based optimization.

How does Cheetah enable machine learning workflows for accelerators?

It supports fast data collection and integrates with ML tooling, demonstrated through reinforcement learning training, gradient-based beamline tuning, gradient-based system identification, physics-informed Bayesian optimization priors, and surrogate modeling for space charge effects.

Aug 3, 2026

Upload by Riley