Neural network-based multiobjective optimization algorithm for nonlinear beam dynamics

Abstract

Multiobjective genetic algorithms (MOGAs) have proven to be powerful in solving multiobjective problems in the accelerator field. Nevertheless, for explorative problems that have many variables and local optima, the performance of MOGAs is not always satisfactory, especially when a small population size is used due to practical limitations, e.g., limited computing resources. To deal with this challenge, in this paper an enhanced MOGA, neural network-based MOGA (NBMOGA), is proposed. In this method, the data produced with the standard MOGA are used to train a neural network. The neural network is fast to produce a large pool of objective function estimates, with sufficiently high accuracy. A subset of the most competitive estimates is selected to form a population (matching MOGA population size), which is then evaluated with the MOGA evaluator. By taking three classic multiobjective problems as examples, we demonstrate that the proposed method promises a faster convergence and a higher degree of diversity than that available with the standard MOGA and other three optimization methods that have been applied in the accelerator field, i.e., the multiobjective particle swarm optimization (MOPSO), the combination of MOPSO and MOGA, and the clustering enhanced MOGA. And then this method is applied to a timeconsuming optimization problem, the dynamic aperture and Touschek lifetime optimization of the high energy photon source. It turns out that, within the same optimization time, a better set of solutions in the objective space can be obtained with the NBMOGA than using other methods. The Touschek lifetime can be improved by about 10% compared with using the standard MOGA, with approximately the same dynamic aperture area. Besides, a higher degree of diversity among solutions is observed with the NBMOGA than using other tested methods.

Design study of APS-U-type hybrid-MBA lattice for mid-energy DLSR

“…Consequently, the design process often relies on identification from a large parameter space, in which the use of efficient algorithms and heavy computation are expected. Therefore, DLSR designs are typically optimized using stochastic optimization algorithms, e.g., the multi-objective genetic algorithm [21][22][23][24][25], multi-objective particle swarm optimization (MOPSO) algorithm [26][27][28][29], or machine-learning-enhanced stochastic optimization algorithms [30][31][32]. In this study, we used MOPSO to optimize the linear dynamics and performed the optimization procedure proposed in Ref.…”

Machine learning enabled fast evaluation of dynamic aperture for storage ring accelerators

“…To reduce the computational cost, machine learning (ML) techniques have recently inspired a surge of applications to DA evaluation, especially in the time-consuming DA optimization studies [29][30][31][32][33]. These proposals focus mainly on learning the map between the magnetic lattice settings of the storage ring (e.g., the strengths of magnets) and the corresponding DA size.…”

Machine learning-based direct solver for one-to-many problems on temporal shaping of relativistic electron beams

“…Particle accelerators, as a collection of multiple complex physical subsystems, has profited from ML since 1980s [33]. In recent years, ML has attracted increasing interests of accelerator experts as a powerful tool to re-veal the complicated correlations between various accelerator parameters [34][35][36][37][38][39][40][41]. Most ML applications in the accelerator field are based on using fast supervised ML surrogate models to predict complex accelerator parameters by learning from the previously existed simulating or experimental data without further simulation or experiment.…”