Speaker
Description
In many advanced accelerator applications, average power is limited by beam loss. The rate and location of losses due to Touschek scattering and residual gas scattering can be accurately computed with knowledge of the lattice. However, these calculations are expensive: involving Monte-Carlo methods where each of the millions of samples is tracked through the machine to the location of its eventual demise. This expense has prohibited the use of loss modeling in an operational or diagnostic setting. In this talk, we introduce a new surrogate model capable of predicting the location of beam loss from its initial scattering coordinates. In experiments on our example lattice, the model achieves ~1 m accuracy at speeds two to three orders of magnitude faster than conventional tracking through the lattice. For many practitioners, the full cost of dataset generation, model training, and inference is less than that of the equivalent classical method through a combination of sharing the model across scattering processes and the sample inefficiency of Monte-Carlo techniques. This work brings loss modeling into the control room - turning a day-long computation into one that is completed in near real time and can be updated interactively with actual and proposed changes to the beam and scattering parameters.
| Working group | WG5 |
|---|