Design and optimization of diffraction-limited storage ring lattices based on invertible neural networks

Purpose In the field of accelerators, identifying optimized operational points and ideal solutions for complex systems remains a significant challenge due to the large number of parameters and intricate nonlinear dynamics involved.
Methods In this study, we present two diffraction-limited storage ring (DLSR) lattice models based on deep and invertible neural networks (INNs), each incorporating both forward and inverse models. These models play a crucial role in enhancing multi-objective evolutionary algorithms (EAs) and expanding the set of viable solutions.
Results and Conclusion We evaluate the accuracy of both the forward and inverse models of two lattice configurations; the accuracy is as high as 97%. Explore the Pareto-optimal trade-offs between emittance and dynamic aperture. INNs were applied in DLSR lattices for the first time, and lattices with natural emittances of 114 pm·rad and 23 pm·rad were obtained at energies of 2 GeV and 6 GeV, respectively, both with reasonable dynamic aperture. Additionally, the use of invertible neural networks significantly reduces computational costs and time requirements. This study provides a valuable reference for future research in multi-objective optimization for lattice design.