Automatic Control of Linear Particle Accelerators with Machine Learning Methods

Particle accelerators are among the most complex physical systems globally, with their design and operation posing significant challenges due to versatile operation modes, stringent beam quality requirements, and the demand for high availability. This dissertation provides a comprehensive overview of integrating machine learning (ML) techniques throughout the life cycle of linear particle accelerators, from design to operation.
In the simulation, the neural networks (NNs) are used as surrogate models for fast, high-quality predictions in place of computationally expensive physics simulations and virtual diagnostics for real-time, non-destructive measurements. In the design phase, parallel Bayesian optimization (BO) is introduced for more efficient parameter optimization. A fast-executing backward-differentiable beam dynamics simulation code Cheetah is developed.
Some of its application cases are highlighted in this dissertation, including gradient-based simulated optimization and support for other algorithms.
This dissertation further explored the use of convolutional neural networks (CNNs) for control of spatial light modulators for laser pulse shaping, with results demonstrated at the FLUTE accelerator. ... mehrSuch fine-grained photo-injector laser pulse shaping is expected to allow tailored generation of electron bunches and increase the accelerator’s dynamic range.
For the online tuning of the accelerator, the performance of BO and reinforcement learning (RL) is compared for online accelerator tuning, with results showing that BO is a turn-key tuning solution and RL has superior performance at the cost of increased upfront engineering effort. Several proposed techniques for building robust and generalizable ML-based controllers across different accelerators are discussed, including domain randomization, meta-RL, and GP-MPC which combines the strengths of BO and RL for beam trajectory tuning tasks.
These advancements highlight the potential of ML-driven solutions in improving the accelerator design, tuning, and control. These preliminary studies pave the way for the development and operation of more efficient and reliable accelerators in the future.