Application of deep learning to soft x-ray tomography at EAST

Abstract

Abstract Soft x-ray (SXR) cameras in a tokamak are limited spatially by ports of the vacuum vessel, and SXR tomography (SXT) technology is developed for reconstructing a two-dimensional SXR profile. However, traditional SXT is time-consuming and has difficulty achieving abundant and quick reconstructions for a tokamak. Based on experimental SXR data and Fourier–Bessel SXT codes at the EAST tokamak, three typical neural networks are built and trained. All the trained neural networks complete reconstruction within several milliseconds on a personal computer and succeed in constraining the SXR profile to match most of the data. In particular, the best-performing fully convolutional neural network provides SXR reconstruction images on the 2D evolution of a sawtooth, and shows its generalization. In the future, it is possible to provide an outstanding deep learning substitute to give abundant and quick SXT images instead of traditional SXT, after training for a few days.