JACoW logo

Joint Accelerator Conferences Website

The Joint Accelerator Conferences Website (JACoW) is an international collaboration that publishes the proceedings of accelerator conferences held around the world.


BiBTeX citation export for TUPAB052: Current Study of Applying Machine Learning to Accelerator Physics at IHEP

@inproceedings{wan:ipac2021-tupab052,
  author       = {J. Wan and Y. Jiao},
  title        = {{Current Study of Applying Machine Learning to Accelerator Physics at IHEP}},
  booktitle    = {Proc. IPAC'21},
  pages        = {1477--1480},
  eid          = {TUPAB052},
  language     = {english},
  keywords     = {network, electron, lattice, target, photon},
  venue        = {Campinas, SP, Brazil},
  series       = {International Particle Accelerator Conference},
  number       = {12},
  publisher    = {JACoW Publishing, Geneva, Switzerland},
  month        = {08},
  year         = {2021},
  issn         = {2673-5490},
  isbn         = {978-3-95450-214-1},
  doi          = {10.18429/JACoW-IPAC2021-TUPAB052},
  url          = {https://jacow.org/ipac2021/papers/tupab052.pdf},
  note         = {https://doi.org/10.18429/JACoW-IPAC2021-TUPAB052},
  abstract     = {{In recent years, machine learning (ML) has attracted increasing interest among the accelerator field. As a complex collection of multiple physical subsystems, the design and operation of an accelerator can be very nonlinear and complicated, while ML is taken as a powerful tool to solve such nonlinear and complicated problems. In this study, we report on several successful applications of ML to accelerator physics at IHEP. The nonlinear dynamics optimization of the High Energy Photon Source (HEPS) that is a 4th-generation light source is a challenging topic. In this optimization, we use a ML surrogate model to fast select the potentially competitive solutions for a multiobjective genetic algorithm that can significantly improve the convergence rate and the diversity among obtained solutions. Besides, we also tried to apply a generative adversarial net to solve one-to-many problems of longitudinal beam current profile shaping. Unlike most supervised machine learning methods than cannot learn one-to-many maps, the generative adversarial net-based method is able to predict multiple solutions instead of one for a 4-dipole chicane to realize several desired custom current profiles.}},
}