Fast estimation of ion temperature from EAST charge exchange recombination spectroscopy using neural network

Abstract

Ion temperature, as one of the most critical plasma parameters, can be diagnosed by charge exchange recombination spectroscopy (CXRS). Iterative least-squares fitting is conventionally used to analyze CXRS spectra to identify the active charge exchange component, which is the result of local interaction between impurity ions with a neutral beam. Due to the limit of the time consumption of the conventional approach (∼100 ms per frame), the Experimental Advanced Superconducting Tokamak CXRS data is now analyzed in-between shots. To explore the feasibility of real-time measurement, neural networks are introduced to perform fast estimation of ion temperature. Based on the same four-layer neural network architecture, two neural networks are trained for two central chords according to the ion temperature data acquired from the conventional method. Using the TensorFlow framework, the training procedures are performed by an error back-propagation algorithm with the regularization via the weight decay method. Good agreement in the deduced ion temperature is shown for the neural networks and the conventional approach, while the data processing time is reduced by 3 orders of magnitude (∼0.1 ms per frame) by using the neural networks.