Gaussian Process Regression and Bayesian Optimization for a 90 MeV Laser-Plasma Injector for the cSTART Storage Ring

Laser-plasma accelerators (LPAs) generate ultrashort high intensity electron bunches from a compact source size. At the Karlsruhe Institute of Technology (KIT), we will use an LPA as one of the injectors for the compact, high-acceptance, non-equilibrium storage ring cSTART.
The LPA injector will be based on an ionization trapping scheme in combination with a tailored plasma density profile to produce an electron beam with small energy spread that maximizes the charge at our target energy, which is at (for LPAs) comparably low energies of 50-90 MeV. Moreover, the LPA injector must produce controlled electron beams with a high shot-to-shot stability and avoid high-energy runaway electrons. These goals can be achieved largely by the detailed design of the plasma density profile and the laser pulse parameters.
In an LPA, small changes across the high-dimensional parameter space can have an outsized influence on overall performance. To handle this challenge, we perform particle-in cell (PIC) simulations and use machine-learning driven approach using Gaussian Process Regression (GPR) and Bayesian Optimization (BO). This procedure allows us to both optimize our gas target design and characterize the effects of the interaction parameters, giving us a functional LPA with a simple tuning mechanism.