Convective heat loads onto the plasma facing components of magnetic confinement devices contain information about edge magnetic field properties which are not yet fully exploited. Machine learning approaches are a promising technique to automatically extract information about such properties from heat load images. In this study, we present the successful reconstruction of proxies for two independent, important edge magnetic field properties given simulated heat load images on the Wendelstein 7-X divertor target plates. Six different artificial neural network architectures from shallow and simple feed-forward fully-connected neural network to deep Inception ResNets with 24 223 to 804 804 free parameters are investigated. The relative reconstruction error is between and with calculation times on the order of milliseconds. A competing benchmark method without machine learning reaches slightly smaller errors but exceeds the calculation time by three orders of magnitude. The experiments demonstrate that machine learning is also a powerful tool in this particular field of nuclear fusion research and deep convolutional neural networks are identified as favorable algorithms for the stated problem. The findings of this paper build a basis for future real time discharge optimization and control by means of machine learning methods.
Real-time equilibrium reconstruction by multi-task learning neural network based on HL-3 tokamak