Real-time multispectral imaging plasma edge analysis through machine learning on the TCV tokamak

Quantitative multispectral imaging enables real-time and offline fusion plasma analysis. Real-time feature recognition algorithms extract plasma edge locations, e.g., emission fronts associated with specific processes or electron temperature values. Offline analysis reconstructs poloidal emissivity profiles of three hydrogenic and three helium lines, inferring plasma parameters and reaction rates. While offline analysis avoids proxies and viewing assumptions, its computational cost prohibits its use in real-time.

We apply machine learning to tomographic reconstruction and plasma parameter inference. For reconstruction, a model-aware architecture uses backprojection followed by a U-Net encoder-decoder network. Trained on phantoms with forward-modelled camera images including specular and lambertian reflections, it yields lower errors than classical algorithms while reducing artefacts. For parameter inference, a deep neural network reconstructs plasma parameters from six reconstructed emissivities. Training data combines collisional radiative models for helium (including singlet-triplet mixing, excluding recombination) and hydrogen (including excitation, recombination, plasma-molecule interactions).

These networks enable real-time 2D reconstructions of electron temperature and density, atomic densities, and hydrogenic reaction rates at 100 Hz on a single Nvidia A100 GPU, allowing for physics-based control applications, and rapid post-discharge analysis.