Linear ion accelerators are widely applicable tools for scientific research and high technology industry. The broad scope of applicability demands a wide range of beam and accelerator operating regimes that entail careful design for desired performance. Optimally meeting targeted beam properties rely on detailed analysis of the interplay between the charged particle beam and material structures that are responsible for producing the electric and magnetic fields used to guide, focus, and accelerate the beam. This interplay is captured in the study of beam dynamics where the transverse and longitudinal motion of beam of charged particles is studied, to the extent possible, separately under reduced models to gain intuition about the system and better form optimal strategies. The integrated system performance is then examined with first principle particle codes that simulate the full 3D dynamics with combined and space charge effects to verify performance. Ideally, this is carried out in source to target simulations.This thesis mainly discusses the theoretical and computational work done to support thedevelopment of compact, cost effective, multi-beam linear ion accelerator systems. Part I provides a brief description of the theoretical concepts needed to better understand particle dynamics in radio frequency linear accelerators. Part II discusses efforts made to optimize conductor geometries used to focus and accelerate beam arrays, and reduced models developed to analyze and better optimize the longitudinal and transverse dynamics in isolation. These reduced models are used to formulate improved systems that are then self-consistently simulated using a particle-in-cell accelerator code called Warp to model full 3D effects including non-linear and space charge effects to better evaluate system performance. Ideas for improved future systems building off the developed computational tool kit are presented. Lastly, Part III details a limited scope project that applied a robust auto tuning algorithm to an existing ion induction accelerator NDCX-II at Lawrence Berkeley National Laboratory. The NDCX-II system has complicated longitudinal pulse compressions that are difficult to optimize for the final pulse durations, spot size, and flux on target. Once found, optimal settings will drift in time as due to various non-static changes (lab conditions such as temperature, etc.). The algorithm was used both as a case study in finding optimum settings for the accelerator, and as a way to sample the accelerator parameter space in order to collect data and build a neural network to act as a surrogate model. With a tuning algorithm, the system can rapidly be brought back to the most-current design optimum and the surrogate model can be used to efficiently probe the parameter space for new optimum, compare performance and identify faults, and more. The tuning algorithm and data cycle is detailed. A neural network is created to predict accumulated charge on target with limited experimental data—approximately 400 experimental runs.
Read
