Towards an integrated plasma perception system for stabilized Z-pinch experiments at ZEI

Modern plasma experiments can generate copious amounts of data, but the problem of how to leverage data quantity to improve inference quality remains unsolved. The inference task can be generalized to finding an unknown vector-valued function of space and time ΠΛ(x,t) = [ρ, v, T, B,…](x,t) that carries all the relevant information about the plasma (density, velocity, etc.). Most often, ΠΛ has to be inferred from indirect diagnostic measurements like chord integrated spectroscopy, resulting in ill-posed inverse problems in the absence of regularization constraints. Most notably, these constraints must include the laws of physics in PDE form. The ZEI integrated plasma perception system uses deep neural networks to model ΠΛ(x,t). By comparing synthetic diagnostic signals computed from ΠΛ to actual diagnostic data, the ΠΛ network can be trained to become the solution to the inverse problem. In addition, the error-free differentiability of neural nets allows for straightforward application of PDE constraints. Experiments with phantoms, bench setups, and initial plasma data at ZEI are presented.