Reconstructing the dynamics of non-linear, kinetic plasma dynamics from partial measurements using physics-informed machine learning

Accurately diagnosing and characterizing plasma dynamics in laboratory experiments is essential for advancing basic plasma science and technology. Modern plasma physics experiments that leverage high repetition-rate lasers and pulsed power systems are providing unprecedented amounts of highly-resolved spatiotemporal plasma measurements, but the development of computational tools that can harness the full potential of such data are still lacking. In this work, we show that physics-informed neural networks (PINNs) can be used to combine partial spatiotemporal measurements of plasma dynamics with fundamental plasma physics equations to reconstruct physically-consistent plasma quantities that were not measured in experiment. We illustrate this approach on PIC simulation data of the nonlinear and kinetic dynamics of electrostatic streaming instabilities. We show that PINNs can reconstruct nonlinear kinetic plasma dynamics from sparse measurement data, and we characterize how the reconstruction accuracy depends on the amount of measurement data and noise level. Finally, we discuss how PINNs can be used to guide and optimize data collection in experiments.