Bayesian optimization of sheath field accelerated ion beams via high-repetition-rate feedback to spectral dispersion

We report analysis of a Bayesian-optimized target normal sheath acceleration (TNSA) experiment at the ELI-Beamlines facility, where the spectral dispersion (and subsequent on-target pulse shape) of the L3-HAPLS laser was modified using outputs from a multivariate Bayesian optimizer trained on data from a Proton Beam Imaging Energy Spectrometer (PROBIES) [1,2]. In addition to the spectrally-resolved proton beam spatial profile provided by PROBIES, proton time-of-flight and Thomson parabola spectrometer diagnostics were run for cross-calibration [3] alongside other diagnostics.



Around 3500 shots onto 10 µm Cu foil were taken during the beam time in batches of 120 shots, retraining the surrogate model used by the optimizer between batches. With total ion yield as the optimization target, a strong trend towards negative offsets in group delay dispersion (to the available limit of -1000 fs²) was consistently observed, while positive GDD offsets resulted in worse performance than nominal best compression (~30 fs, ~10²¹ W/cm², a₀ ~ 20). Models trained on large ensembles of PIC simulations show good qualitative agreement with the data.

[1] D. Mariscal et al., Plasma Physics and Controlled Fusion 63, 114003 (2021)

[2] D. Mariscal et al., Rev. Sci. Instrum. 94(2) (2023)

[3] I. Rodger et al., Rev. Sci. Instrum. (In review)