Developing Neural Networks for Accelerator Beam Phase Space Tomography

High-energy particle accelerators require an in-depth understanding of the beam properties to
operate effectively. The characteristics of beams have been studied using a variety of methods,
but as accelerator machines become increasingly advanced, more sophisticated techniques are
required to understand their behaviour. Modern diagnostic and analysis methods often demand
the processing of large amounts of data, which involves significant computational power and
run-time to complete a given analysis, limiting the applicability of some traditional techniques.
This PhD project aims to develop artificial intelligence and machine learning methods for
characterising particle accelerators and beam behaviour. These methods offer the possibility to
greatly improve accelerator tuning and operation efficiency by processing the large data sets
required with a fraction of the computing resources.
As an example, machine learning has been applied to measure the charge distribution within
bunches of high-energy electrons using CLARA, the Compact Linear Accelerator for Research
and Applications at Daresbury Laboratory [1]. The use of machine learning for phase space
tomography allows the reconstruction of the phase space distribution of the electron bunches by
training neural networks to recover the distribution from beam images.
The computing requirements for traditional phase space tomography techniques increase rapidly
with increasing dimensionality of the phase space. The storage of a D-dimensional distribution in
an array with dimension length N requires a data set structure of ND values, and the processing of
this data to reconstruct the phase space can require considerable memory resources. The amount
of memory required to store a data set representing four-dimensional phase-space distributions
with a moderate resolution of N < 100 pixels can be a several gigabytes.
Image compression techniques allow for the reduction of the processing time and data storage
requirements for phase space tomography. In contrast to conventional tomography methods,
machine learning techniques enable direct tomographic analysis of compressed data, greatly
assisting the efficiency of the phase space analysis.
Once the beam properties have been characterised, the accelerator is usually tuned to achieve
specified performance goals. The application of traditional tuning methods based on modelling of
the particle accelerator are often effective, but these can be difficult and time-consuming to apply.
Neural networks are trained using large data sets generated from detailed and computationally
intensive models, but they quickly generate results once trained, offering the possibility of
developing more accessible tuning and optimisation tools for the control room.
While understanding the overall behaviour of advanced accelerators depends on knowledge of
beam properties such as the phase space distribution, a range of other applications of machine
learning can also be of potential benefit for accelerator research and development.
This PhD project will explore the potential of machine learning to facilitate optimisation of
accelerator performance and improve diagnostic techniques. There will be a focus on CLARA,
but generalisation to different types of accelerators will also be considered.
There are several essential components to the project other than training neural networks and
modelling accelerator dynamics, including the validation of methods based on machine learning
in experimental tests, assessment of the feasibility of implementing these novel techniques in
the control room, and evaluation of the advantages and disadvantages of machine learning in
comparison to conventional methods.
References
[1] A. Wolski, M. A. Johnson, M. King, B. L. Militsyn, and P. H. Williams (2022). "Transverse
phase space tomography in the CLARA accelerator test facility using image compression
and machine learning." arXiv:2209.00814