Self-supervised deep learning for intense charged particle beam dynamics with hard physics constraints

The dynamics of intense relativistic charged particle beams in can be dominated by complex collective effects in which the behavior of every individual particle depends on the entire ensemble. At low energy (< 5 MeV) high charge beams are dominated by collective space charged forces. At high energy many particle accelerators (such as the beam-driven plasma wakefield accelerator FACET-II) compress charged particle beams to very high intensity with individual bunches having up to 2 nC or charge compressed to a length of only a fraction of a micron. At such compression a relativistic beam's length in time (when moving near the speed of light) is only ~1 femtosecond. Such a compressed beam has a peak current as high as 2000 kA. In this case space charge forces can be important even for highly relativistic beams. The simulation of the dynamics of such beams can be extremely computationally expensive based on O(n^2) individual particle-to-particle interactions, where n=1.25x10^10 for a 2 nC bunch. In this talk we present recent results on developing deep neural network-based self-supervised deep learning for such intense charged particle beam dynamics with guaranteed hard physics constraints. We demonstrate how this method, when coupled with more traditional physics-based solves, has the potential to greatly speed up calculations for simulating the dynamics of intense charged particle beams.