20-dimensional surrogate-assisted Bayesian optimization of laser-driven proton beams

Abstract

Laser-driven proton acceleration, as obtained by the interaction of a high-intensity laser with matter, is a promising technique for generating high-quality proton beams. One of the main challenges is increasing the maximum proton energy. Here, we demonstrate a 70% increase in the maximum energy of laser-driven protons by optimizing the wavefront of the intense laser using machine learning: This was accomplished through adaptive control of a deformable mirror (DM) using a multi-step Random Forest surrogate-assisted Bayesian optimization approach. Starting from zeroed DM actuator voltages, our method identified an optimal configuration using 20 out of 48 actuators, requiring fewer than 150 experimental data samples. Our method surpassed conventional wavefront correction by 24%, which typically minimizes aberrations to converge toward a flat wavefront by leveraging real-time feedback from a wavefront sensor. This data-driven method integrating advanced wavefront control challenges the preference for correcting aberrations to achieve a flatter wavefront in laser-driven ion acceleration. We also propose a strategy for optimizing short focal length ion accelerators at facilities where measuring the wavefront at nominal full laser power is not implemented.