Faraday angle prediction based on 1-D CNN on the J-TEXT tokamak

Abstract

The Faraday angle is an indispensable physical parameter to evaluate and reconstruct the current density profile which is imperative for advanced tokamak operation and physics study. The POLARIS system has been implemented on J-TEXT since 2012 for the measurement of the Faraday angle and this system occasionally encounters operational failures that result in the inability to provide measurements. So a deep learning model is trained to predict the Faraday angle when the POLARIS is unable to provide the measurement. The model is primarily based on a 1-D convolutional neural network (CNN) and is equipped with multiple convolutional paths that can accommodate various lengths of diagnostic data. The data of 100 shots are used to train the 1-D CNN model, and 30 shots are used to optimize the hyperparameters and select the model which performs best during the training process. The results on a test set of 30 shots show that the model obtains a satisfactory fitting effect.