Optimisation of linear accelerator performance for single-pass free-electron laser operation

This project is part of a collaboration betweenMonash University, the Aus-
tralian Synchrotron (AS), the new FERMI@Elettra project, and the Linac Co-
herent Light Source (LCLS) at the SLAC National Accelerator Laboratory.
The thesis investigates the use of Artificial Intelligence systems and their
applicability to machine optimisation and control for linear accelerators,
and in particular Free Electron Lasers (FEL). This research is motivated by
the need to develop adaptive systems for beam tuning and stabilisation,
in order to meet the increasingly stringent requirements of new generation
light sources.

The thesis begins with the simulation of a feedback system for the
FERMI@Elettra Linac, based on the Proportional - Integral - Differential
(PID) scheme developed for the LCLS. To facilitate these simulations, aMat-
lab Graphical User Interface was built in order to incorporate various con-
trol parameters, including possible actuators, observable variables and per-
turbations. These simulations highlight the difficulties encountered with
PID control, which necessitates a more sophisticated approach to the con-
trol of the linear accelerator.

To address the intrinsic limitations of conventional PID control, a com-
bination of a feedforward - feedback system was investigated. The feedfor-
ward component uses a neural network (NNet), which provides a prediction
of the perturbation in the electron beam parameters based on fluctuations
in the voltage and phase of the klystrons. The feedback component con-
sists of a simple PID algorithm, used to compensate for potential inaccura-
cies of the feedforward correction. Experimental results performed at the AS show the viability of the system, by demonstrating the successful con-
trol of the energy at the end of the Linac. The experiments carried out at the
LCLS show the applicability of the system to a multi-variable system with
the simultaneous control of the energy and bunch length. Although these
results demonstrate the ability of the NNet predictor to compensate for the
deficiencies of the PID algorithm, further refinements of the technique are
required to produce a system that can adapt to changes in machine param-
eters and jitter conditions.

To correct the remaining deficiencies of the combined feedforward -
feedback control system, we consider an intelligent system capable of self-
learning. In this scenario the control system is treated from the perspective
of an optimisation problem and a novel optimisation tool is designed, us-
ing state of the art developments in video games. The key principle is to
exploit similarities between the navigation of a game agent in a battlefield
and the navigation of an "optimisation agent" in a search space. This novel
approach was tested using simulations and experiments conducted at the
AS and on the FERMI@Elettra Linac. The experiments conducted at the
AS showed the system’s ability to simultaneously optimise the beam energy
spread and transmission (i.e. the percentage of particles transmitted from
the start to the end of the accelerator). We have demonstrated an increase
in the transmission from 90% to 97% and a decrease in the energy spread of
the beam from 1.04% to 0.91%. Control experiments performed at the new
FERMI@Elettra FEL are also reported, which highlight the adaptability of
the system for beam-based control, in the case where a static perturbation
is applied to the klystron phase. These results show that NNets can be suc-
cessfully exploited to build an optimisation tool that can self-learn from its
interaction with the machine and operate a simple control task. Our results
indicate that this optimisation tool can be used for the stabilisation of the
electron beam parameters when it is subject to time dependent perturba-
tions.
The thesis concludes with suggestions for future work. This includes
the adaptation of the optimisation tool to N-dimensional search spaces,
and the development of a novel control system which merges the NNet pre-
dictor with the structure of the NNet used for optimisation. By combining
these two structures, it is anticipated that the resulting NNet will have the
ability to correct time dependent perturbations, while self-adapting its re-
sponse when machine parameters and jitter conditions change.

本项目隶属于莫纳什大学、澳大利亚同步加速器（Australian Synchrotron, AS）、新兴的费米@埃尔特拉（FERMI@Elettra）项目，以及SLAC国家加速器实验室旗下直线加速器相干光源（Linac Coherent Light Source, LCLS）的合作研究计划。
本论文探讨了人工智能系统在直线加速器，尤其是自由电子激光器（Free Electron Lasers, FEL）的机器优化与控制中的应用。本研究的初衷是开发适用于束流调谐与稳定的自适应系统，以满足新一代光源日益严苛的性能要求。
论文首先基于为LCLS开发的比例-积分-微分（Proportional-Integral-Differential, PID）控制方案，对FERMI@Elettra直线加速器的反馈系统开展仿真研究。为支撑此类仿真工作，研究人员搭建了Matlab图形用户界面，用于集成各类控制参数，包括可选执行器、可观测量与扰动源。仿真结果凸显了PID控制的固有缺陷，表明直线加速器的控制亟需更先进的解决方案。
为克服传统PID控制的固有局限，本研究探索了前馈-反馈复合控制系统。其中前馈模块采用神经网络（Neural Network, NNet），可基于速调管的电压与相位波动预测电子束参数的扰动；反馈模块则采用简易PID算法，用于补偿前馈校正可能存在的误差。在AS开展的实验验证了该系统的可行性：成功实现了直线加速器末端的束流能量控制。在LCLS开展的实验则证明，该系统可应用于多变量系统，实现束流能量与束团长度的同步调控。尽管上述结果证实了NNet预测器可弥补PID算法的不足，但仍需对该技术进行进一步优化，方能构建出可适配机器参数变化与抖动条件的自适应系统。
为进一步修正前馈-反馈复合控制系统的剩余缺陷，本研究考虑采用具备自学习能力的智能系统。本研究将控制系统视作优化问题，并依托电子游戏领域的当前最先进技术，设计了一款新型优化工具。其核心原理在于利用游戏智能体在战场环境中的导航逻辑，与优化智能体在搜索空间中的寻优逻辑之间的相似性。本研究通过仿真以及在AS、FERMI@Elettra直线加速器上开展的实验，对该新型方法进行了验证。在AS开展的实验表明，该系统可同时优化束流能散与传输效率（即粒子从加速器入口到出口的传输百分比）：研究实现了传输效率从90%至97%的提升，以及束流能散从1.04%至0.91%的降低。本文还报告了在新型FERMI@Elettra FEL上开展的控制实验，结果显示当对速调管相位施加静态扰动时，该系统具备基于束流的控制适应性。上述结果证实，可成功利用NNet构建具备自学习能力的优化工具，该工具可通过与加速器的交互自主学习并完成简单控制任务。研究结果表明，当电子束参数受到时变扰动时，该优化工具可用于稳定束流参数。
论文最后对未来研究方向提出了建议，包括将该优化工具适配至N维搜索空间，以及开发融合NNet预测器与优化用NNet结构的新型控制系统。通过整合这两种结构，预期最终得到的神经网络可具备校正时变扰动的能力，同时可在机器参数与抖动条件发生变化时自适应调整其响应。