LBNL, Berkeley, California, USA
The bare lattice optimization for the linear and nonlinear ALS-U storage ring lattice, even without reverse bending, comprises 11 degrees of freedom (DoF) and is therefore a very complex and highly time-consuming process. This design process relies heavily on multi-objective genetic algorithms (MOGA), usually requiring many months of experienced scientists’ time. The main problem lies in having to evaluate numbers of candidate lattices due to the stochastic process of MOGA. Although almost all of these candidates are eventually rejected, they nevertheless require extensive particle tracking to arrive at a Pareto front. We therefore propose a novel Machine Learning (ML) pipeline that nonlinear tracking is replaced by two well-trained neural networks (NNs) to predict dynamic aperture (DA) and momentum aperture (MA) for any lattice candidate. Initial training of these models takes only several minutes on conventional CPUs while predictions are then rendered near instantaneously. We present this novel method and demonstrate the resulting orders of magnitude speedup of the ML-enhanced MOGA process on a 2-DoF problem as well as first results on a more complex 11-DoF problem.
