Time-dependent analysis of edge plasma turbulence via deep learning from partial observations

We demonstrate that a physics-informed multi-network deep learning architecture constrained by partial differential equations can accurately learn turbulent fields consistent with drift-reduced Braginskii theory from just partial observations of electron pressure in contrast with conventional analytic equilibrium methods. This framework further enables the first ever direct quantitative comparisons of electron pressure and electric field fluctuations in nonlinear global electromagnetic gyrokinetic simulations and electrostatic two-fluid theory. Accordingly, we quantitatively explore the concomitant response that exists between the fluctuating electron pressure and electric potential which constitutes one of the key relationships demarcating a plasma turbulence model. The methods outlined can be readily adapted to the study of magnetized quasineutral plasmas in advanced geometries and presents broad implications for the validation of reduced plasma turbulence models in experimental and astrophysical settings. In particular, applications of our deep learning framework to tokamak diagnostics for time-dependent analysis and interpretation of edge turbulent fluctuations measured by gas puff imaging will be considered.