Accelerator beam phase space tomography using machine learning to account for variations in beamline components

Aug 4, 2026

Upload by Asher

A technique is presented for reconstructing the four-dimensional transverse phase space of an accelerator beam along a beamline while explicitly accounting for unknown errors in the strengths of magnets used during data collection. Machine learning enables rapid recovery of the phase-space distribution and simultaneously provides estimates of magnet errors. The method is validated with experimental measurements from CLARA, an accelerator test facility at Daresbury Laboratory.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages32

Reading Time1 h 21 min

AccessFree to download

1 Introduction: machine learning for accelerator beam phase space tomography

2 Tomography with extended latent space to account for errors on accelerator components

3 Implementation of the extended latent space technique: an example

3.1 Training data

3.2 Sinogram autoencoder

3.3 Extended encoder

3.4 Phase space decoder

4 Experimental demonstration

5 Conclusions

arXiv :2405 . 10028v1 [physics .acc-ph] 16 May 2024 Prepared for submission to JINST Accelerator beam phase space tomography using machine learning to account for variations in beamline components A. Wolski,􀀰,􀀲, 1 D. Botelho,􀀰,􀀲 D. Dunning,􀀱,􀀲 A. E. Pollard􀀱,􀀲 􀀰 Department of Physics, University of Liverpool, Oxford Street, Liverpool L69 7ZE, UK 􀀱 ASTeC, STFC Daresbury Laboratory, Daresbury, WA4 4AD, UK 􀀲 The Cockcroft Institute, Sci-Tech Daresbury, Keckwick Lane, Daresbury, WA4 4AD, UK E-mail: [a.wolski@liverpool.ac.uk](a.wolski@liverpool.ac.uk) Abstract: We describe a technique for reconstruction of the four-dimensional transverse phase space of a beam in an accelerator beamline, taking into account the presence of unknown errorson the strengths of magnets used in the data collection. Use of machine learning allows rapid reconstruction of the phase-space distribution while at the same time providing estimates of the magnet errors. The technique is demonstrated using experimental data from CLARA, an accelerator test facility at Daresbury Laboratory. Keywords: Beam dynamics; Beam optics; Analysis and statistical methods; Data reduction methods 1Corresponding author. Contents 1 Introduction: machine learning for accelerator beam phase space tomography 1 2 Tomography with extended latent space to account for errors on accelerator compo nents 4 3 Implementation of the extended latent space technique: an example 10 3.1 Training data 10 3.2 Sinogram autoencoder 12 3.3 Extended encoder 13 3.4 Phase space decoder 17 4 Experimental demonstration 19 5 Conclusions 27 1 Introduction: machine learning for accelerator beam phase space tomography Many modern accelerators rely on the production of high-quality particle beams to reach their performance specifications. Facilities such as X-ray free-electron lasers, for example, have demanding requirements for electron beam transverse and longitudinal emittance [1–4] . Commissioning, tuning and effective operation of many accelerators require the capability to make rapid and reliable measurements of beam parameters at different points along a beamline, often starting at the particle source. In the case of the transverse beam emittance, quadrupole scans provide a standard measurement technique [5, 6]: the emittance is found from the dependence of the beam size (measured using, for example, a wire scanner or imaging screen) on the strength of an upstream quadrupole magnet. The use of tomographic methods [7–14] allows the transverse phase-space distribution of the beam to be reconstructed from the screen images, yielding not just the emittance but also the optics functions describing how the distribution changes along the beamline. In addition, phase space tomography can provide information on coupling between transverse planes [15, 16] and on the detailed charge distribution within bunches. With the use of RF transverse deflecting structures, it is possible to make detailed measurements of the longitudinal phase space [17, 18] . Combining quadrupole scans with a transverse deflecting cavity makes it possible to determine the five or six-dimensional phase-space beam distribution [19–22] . Despite the fact that phase space tomography based on quadrupole scans is a well-established technique, there remain several significant challenges in its application, especially for low-emittance beams. Data collection and processing can be time-consuming, especially for higher-dimensional phase space reconstruction. Traditional tomography algorithms can be prone to artefacts in thereconstruction, especially when the range or number of projection angles is limited (as is often the case in beam phase space tomography) [23, 24]: although there are techniques that can be used to reduce the appearance of reconstruction artefacts (see, for example,[25]), ultimately the fidelity of the reconstruction depends on collecting screen images with a sufficient number of diffe

FAQs

What problem does the technique address in beam phase space tomography?

It reconstructs transverse phase space while accounting for unknown errors in magnet strengths that affect measurements during data collection.

How does machine learning contribute to the reconstruction and error estimation?

It enables fast reconstruction of the phase-space distribution and provides estimates of the magnet errors at the same time.

What experimental facility was used to demonstrate the approach?

The technique is demonstrated using experimental data from CLARA, an accelerator test facility at Daresbury Laboratory.

Aug 4, 2026

Upload by Asher