Accelerating predictive modeling of laser-driven ion acceleration with deep learning

Fully kinetic simulations are commonly used to model the physics of intense laser-plasma interactions and ion acceleration. These methods offer an accurate description of the plasma microphysics, but at significant computational cost, which has limited the reach of 3D simulations and the ability to develop predictive simulations of laser-driven secondary sources. In this work, we explore deep learning-based methods for accelerating the modeling of nonlinear, many-body systems, with a focus on laser-driven ion acceleration. We aim to learn efficient, low-dimensional representations of the phase-space structure of the plasma and their associated evolution operators. In addition, we aim to incorporate fundamental symmetries in the machine learned representations to ensure physical consistency and generalizability. We will discuss results obtained based on data from high-fidelity particle-in-cell simulations and show the potential of the new methods for accelerating the modeling of laser-driven ion acceleration systems.