Combining Bayesian Optimization and Neural Network to Optimize the Plasma Wakefield Acceleration

Abstract

The computational cost of finding the optimal design of plasma wakefield acceleration (PWFA) is usually very demanding due to many variables involved. Herein, we have developed a novel framework which combines Bayesian Optimization (BO) with neural network (NN), to replace computationally expensive simulation software and provide a more efficient way for the optimization process. In order to verify this framework, the AWAKE Run 2 experiment at CERN is used as an example. In the framework we constructed, the coefficients of determination (R2) of NN reaches above 0.99, and the time-to-solution reduces to a factor of 18.6. For the first time, BO combined with NN is successfully applied to optimize PWFA and significant improvements have been demonstrated. The framework established here in principle can also be extended to the optimization of other particle accelerations.