Faster HPC Particle-in-Cell Simulations of Hybrid Particle Accelerator Beamlines by In Situ Coupling of Machine Learning Surrogate Models

Particle accelerator simulations are essential to investigating, designing and operating some of the most complex scientific devices ever built. Kinetic simulations of relativistic, charged particle beams and advanced plasma accelerator elements are often performed with high-fidelity particle-in-cell simulations, some of which fill the largest GPU supercomputers. Start-to-end modeling of future particle accelerators includes many beamline elements for transport, control and acceleration, and it is thus desirable to integrate and model the central, advanced accelerator elements fast. Traditionally, analytical and reduced-physics models are used to achieve speedups over full-fidelity first-principles simulations. A new opportunity to complement modeling without approximations is opened up by using the vast data from high-fidelity first-principles simulations and GPU-accelerated computation for surrogate modeling through machine learning. We implement and benchmark such a data-driven modeling approach in the Beam, Plasma & Accelerator Simulation Toolkit (BLAST), creating an open conventional-surrogate simulation framework for hybrid particle accelerator beamlines.