Surrogate model of particle accelerators using encoder–decoder neural networks with physical regularization

Abstract

Accelerator engineering could benefit from faster and higher-quality physics simulations. Machine learning has emerged as a promising tool for developing accelerator simulation programs that are both fast and accurate. In this study, we propose a surrogate model based on encoder–decoder neural networks. We incorporate physical regularization into the loss function, which allows us to integrate prior physical knowledge into the deep learning network. The advantage of this regularization is that it ensures the results are more consistent with the underlying physical laws. The method was applied to beam dynamics modeling of the medium energy beam transport (MEBT) section in the China Accelerator Facility for Superheavy Elements (CAFe II). The final results indicate that after training, the network maintains a mismatch, emittance difference, and transmission efficiency of approximately 0.01. Our scheme has been demonstrated to be effective in the simulation of the accelerator’s beam dynamics.