Speaker
Description
The study of Dynamic Aperture (DA) plays a crucial role in understanding non-linear beam dynamics in circular accelerators. The DA defines the phase-space region where particles' motion remains bounded over a finite number of turns. It is affected by various elements such as the regular magnetic lattice, magnetic field imperfections, beam-beam effects, electron clouds, and other nonlinear phenomena. Investigating the DA offers valuable insights into beam loss evolution, which is vital for the design of future accelerators like the Future Circular Collider.
Traditionally, numerical evaluation of the DA involves computationally-intensive simulations of initial conditions distributed in phase space over a realistic time interval. In this work, we propose a novel approach utilizing two deep neural networks: the first network regresses the DA values, while the second network estimates the error associated with the DA estimation, leveraging machine parameters.
Through extensive training, our models enable fast and smart sampling. When the estimated error from the second network is within an acceptable range, we utilize the DA value provided by the first network. However, if the estimated error exceeds the threshold, we resort to the conventional simulation approach with tracking simulations, accumulating sufficient samples for subsequent training.
This active learning framework allows for efficient exploration of machine parameters space, reducing computational demands while maintaining accuracy. Our approach demonstrates the potential for accelerating DA simulations and offers a promising avenue for improving the design and tuning of machine parameters for future circular accelerators.
