Machine learning tomography for sparse-view soft x-ray diagnostics in MAST-U

Abstract

Understanding magnetohydrodynamic (MHD) instabilities is crucial for advancing magnetic confinement fusion. Soft x-ray (SXR) tomography enables reconstruction of the spatial structure of these instabilities from one dimensional line integrated measurements, however traditional tomography methods struggle with sparse diagnostic arrays. The MAST-U tokamak has a sparse camera array of 28 non-intersecting lines of sight (LoS), making traditional tomography techniques challenging. We present a machine learning (ML) tomographic method using a 50 layer residual neural network trained on synthetic data simulated by the transport code TRANSP. ML tomography is compared with traditional minimum Fisher information tomography, with second derivative smoothing tomography, and with Gaussian process tomography. The ML technique shows a 23 to 94 times median improvement in the mean square error over traditional 28 LoS reconstructions of synthetic data, at a computational speed which is 15 to 418 times greater than the minimum Fisher and second derivative methods (27 ms on graphical processing unit hardware). Notably, ML tomography with 28 LoS outperforms traditional methods even when provided with three times as much data in the form of 84 intersecting synthetic LoS. Validation on experimental data from MAST-U demonstrates accurate reconstructions of peaked and broad plasma SXR profiles, and the evolution of sawtooth instabilities are visualized with ML tomography. The sub-30 ms reconstruction time enables real-time plasma diagnostics, opening new possibilities for MHD monitoring in fusion devices.