MC6: Beam Instrumentation, Controls, Feedback and Operational Aspects
T35 Machine Learning
Paper Title Page
WEVIR10 Adaptive Feedback Control and Machine Learning for Particle Accelerators 53
 
  • A. Scheinker
    LANL, Los Alamos, New Mexico, USA
 
  The precise control of charged particle beams, such as an electron beam’s longitudinal phase space as well as the maximization of the output power of a free electron laser (FEL), or the minimization of beam loss in accelerators, are challenging tasks. For example, even when all FEL parameter set points are held constant both the beam phase space and the output power have high variance because of the uncertainty and time-variation of thousands of coupled parameters and of the electron distribution coming off of the photo cathode. Similarly, all large accelerators face challenges due to time variation, leading to beam losses and changing behavior even when all accelerator parameters are held fixed. We present recent efforts towards developing machine learning methods along with automatic, model-independent feedback for automatic tuning of charge particle beams in particle accelerators. We present experimental results from the LANSCE linear accelerator at LANL, the EuXFEL, AWAKE at CERN, FACET-II and the LCLS.  
video icon
 
DOI • reference for this paper ※ https://doi.org/10.18429/JACoW-IPAC2020-WEVIR10  
About • paper received ※ 27 May 2020       paper accepted ※ 12 June 2020       issue date ※ 14 June 2020  
Export • reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)  
 
WEVIR12 Machine Learning Techniques for Optics Measurements and Corrections 61
 
  • E. Fol, R. Tomás García
    CERN, Meyrin, Switzerland
  • G. Franchetti
    GSI, Darmstadt, Germany
 
  Recently, various efforts have presented Machine Learning (ML) as a powerful tool for solving accelerator problems. In the LHC a decision tree-based algorithm has been applied to detect erroneous beam position monitors demonstrating successful results in operation. Supervised regression models trained on simulations of LHC optics with quadrupole errors promise to significantly speed-up optics corrections by finding local errors in the interaction regions. The implementation details, results and future plans for these studies will be discussed following a brief introduction to ML concepts and its suitability to different problems in the domain of accelerator physics.  
video icon
 
DOI • reference for this paper ※ https://doi.org/10.18429/JACoW-IPAC2020-WEVIR12  
About • paper received ※ 02 June 2020       paper accepted ※ 12 June 2020       issue date ※ 16 June 2020  
Export • reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)