Deep Learning for Plasma Tomography in Nuclear Fusion

Abstract

Tomography is arguably one of the most representative examples of an inverse problem, where the shape of an object must be reconstructed from its projections over a limited number of lines of sight. The regularization that must be imposed to solve such an ill-posed problem often results in iterative algorithms that are computationally expensive and do not meet the requirements of real-time applications. Deep learning offers a promising approach to perform such reconstruction with sufficient accuracy, while being several orders of magnitude faster, to the point that it becomes possible to use tomography in real-time. In this paper, we give an example of how real-time tomography based on deep learning is being used to reconstruct the plasma radiation profile in a nuclear fusion device. The availability of such profile in real-time allows setting up new alarms in the real-time control system, with a view towards anticipating plasma disruptions.

Tasks

Reproductions