Effective Accelerator Operation Based on Machine Learning at SACLA

Abstract

Machine learning methods have been applied to XFEL optimization, fault prediction, and anomaly detection with the common goal of rationalizing the accelerator operations at SACLA. A Gaussian process optimizer was developed to maximize the XFEL performance, and the spectral brightness was improved by a factor of 1.7 with a new high-resolution inline spectrometer. Further ongoing efforts are also described. As for fault prediction and anomaly detection, three applications were developed: 1) failure prognosis of thyratrons, 2) prediction of recovery time after an interruption of operation, and 3) numerical reading of instruments by image recognition. Although the accuracy of the prototype applications of 1) and 2) are not yet sufficient, they will be improved in future developments. The application 3) has enough accuracy and the readout error of a variable-area flowmeter is a few %.