Machine learning applications for plasma diagnostics

Machine learning (ML) has become an increasingly popular tool in the plasma physics community, especially for plasma control applications such as disruption identification. In addition, ML can play an import role in advanced diagnostic analysis. We will present different applications of neural networks (NN) that highlight some of the potential of these ML applications. First, we will demonstrate a technique using Thomson Scattering to train a NN with multi-energy Soft X-ray measurements, along with other diagnostics, to enable fast measurements of the electron temperature profile. Also, we will present a NN based technique to combine slow, yet accurate, foil bolometry measurements of radiated power with faster diode-based measurements to provide both fast and accurate measurements of P_rad. These are examples of using machine learning to learn the complex relationship between various diagnostics and the desired measurement value. Finally, we will use NN-based classification to identify impurities using spectroscopic measurements. One weakness of NN-based techniques is the difficulty in applying a trained network to other input systems. We will explore the portability of a trained NN by adding a ‘translation’ layer to the input.