Speaker
Description
In accelerator physics, lattice optimization is a foundational step for subsequent work. Traditional optimization methods face problems that balancing several input parameters with conflicting optimization objectives and the prohibitive computational cost of key physics simulations, particularly for Dynamic Aperture (DA). As a part of the China Spallation Neutron Source (CSNS) FFAG project, this study aims to enhance the efficiency and reliability of its multi-objective lattice optimization workflow.
This study focuses on improvements to two critical components of this process. First, we refine the elite clustering process to increase the stability of the genetic algorithm (GA), leading to more robust optimization results. Second, for the time-consuming physics simulations, we have verified the unique suitability of a Transformer-architecture surrogate model for this application. Compared to conventional neural networks, its prediction accuracy is higher.
The combination of the stability-improved genetic algorithm and the high-precision Transformer surrogate model provides a fast set of optimization tools for lattice. It can provide guidance for the choose of lattice parameters in FFAG and significantly accelerate the design and optimization time for the actual engineering project.
