Accelerator tuning method using autoencoder and Bayesian optimization

Abstract

During beam-accelerator operation, a large number of parameters need to be tuned. In recent years, tuning methods based on machine learning have been extensively studied. Bayesian optimization (BO) has attracted considerable attention as an excellent method for accelerator tuning. However, its applicability is limited by the number of parameters that can be tuned. In this study, we propose an optimization method that combines autoencoder and BO to tune a large number of parameters. We verified it using beam transport simulations. We confirmed a higher tuning effect in a shorter time than when using only BO. The proposed method is expected to speed up the accelerator operation and provide comprehensive tuning.