Machine learning controlled laser wakefield acceleration simulations

Abstract

One of the most promising technologies to form the next generation of compact particle accelerators is plasma acceleration. Plasmas have the ability to sustain waves with electric fields that can be three orders of magnitude higher than those in radio frequency (RF) cavities.The ultimate goal of plasma-based acceleration is to produce relativistic, high quality electron and positron bunches for scientific and societal applications. The recent progress has been tremendous but improving beam quality still remains as a grand-challenge in the field.The fundamental aspects and properties of these accelerators are accessible through simplified analytical models, but the self-consistent dynamics of the laser in the plasma can only be captured by numerical simulations. Search for optimised parameters to improve beam quality can be based on systematic parameter scans. However, because numerical calculations can be very computationally intensive, it is important to investigate more efficient techniques to scan over the entire parameter range currently available. In this work, we propose a machine learning approach to optimize this search based on genetic algorithms.Recent experiments have employed genetic algorithms to control plasma based accelerators[1]. Here, instead, we will employ this technique to control the outputs and optimise plasma-based accelerators in particle-in-cell (PIC) simulations. We implemented a genetic algorithm in ZPIC, a fully relativistic PIC educational code[2]. The genetic algorithm is fully automated: it receives an initial set of input parameters, launches several simulations in parallel using MPI, and ends automatically once given convergence criteria are reached. The algorithm can thus take full advantage of large-scale super-computers. We present results from 1D simulations.We focus on plasmas with non-uniform density and lasers with variable longitudinal envelope profiles.