Deep Learning Based Surrogate Model a fast Soft X-ray (SXR) Tomography on HL-2 a Tokamak

Abstract

Tomography is indeed a commonly employed diagnostic technique in fusion campaigns, specifically for determining the shape and position of the plasma. To enhance the accuracy of conventional tomography algorithms, a Bayesian-based non-stationary Gaussian processes tomography as the emission model has been implemented in the soft X-ray diagnostics of the HL-2 A tokamak. However, the Bayesian tomography method is time-consuming and has difficulty achieving quick reconstructions for tokamak. In this work, neural networks have been trained and tested on a large set of sample tomograms based on experimental SXR data and Bayesian tomography method. The trained neural networks can predict the reconstructions of emission profiles accurately, fast, and robustly to noise. In the future, it is possible to easily implement this algorithm on different diagnostics and fusion devices.