Real-Time Machine-Learning Enabled Emission Front Control at DIII-D

A novel method of obtaining emission front locations based on real-time tangential TV (rt-TangTV) images in DIII-D can be used to instantaneously control divertor detachment. The correlation of gas fueling with detachment can be observed in plasma radiation images, where the emission front detaches from the divertor target plate with increased fueling. The rt-TangTV provides a tangential view of C-III radiation, allowing the radiation front location to be determined using tomographic inversion, assuming poloidal symmetry. This process is computationally intensive and may be a bottleneck for the real-time detachment control based on rt-TangTV.

We have developed fully data-driven, regression-based machine-learning methods that can extract the location of the emission front with a cross-sectional accuracy of approximately 1 cm. This approach enables precise real-time control of the emission front. Models are trained on front locations corresponding to rt-TangTV images. It is intended for use at DIII-D to maintain L-mode and rev-B H-mode emission fronts are positioned roughly at a target height between the X-point and target plate, regardless of radial displacement. This control can help sustain good plasma performance with a stable heat flux-controlled scenario.