Reconstruction of poloidal magnetic fluxes on EAST based on neural networks with edge electromagnetic diagnostics

Abstract Accurate reconstruction of tokamak equilibria requires precise measurements of magnetic fields and magnetic flux, which is essential for effective plasma control and optimization. To reconstruct equilibria, code like equilibrium fitting code (EFIT) solves the nonlinear Grad–Shafranov equation. This depends on the traditional algorithms achieved by iteratively minimizing the discrepancy between predicted and measured edge electromagnetic diagnostic signals. The iterative methods face high computational cost due to necessary convergence and sensitivity to initial guesses in the pursuit of equilibrium optimization. We develop an artificial neural network with a weighted focused loss function to reconstruct magnetic equilibria on the EAST tokamak through a supervised learning method. The training set, validation set, and testing set are partitioned randomly from the dataset of poloidal magnetic flux distributions of the EASTexperiments in 2016 and 2017 years. The accuracy of reconstructions is evaluated using a variety of indices, such as the mean squared error (MSE), peak signal-to noise ratio (PSNR), and structural similarity index measure (SSIM), and similarity (S) with Fr´echet distance. The feasibility of the neural network model is verified by comparing it to the offline EFIT results. It is found that the neural network algorithm based on the supervised machine learning method can accurately predict the location of different closed magnetic flux surfaces at a high efficiency. The similarities of the predicted X-point position and last closed magnetic surface are both 98%. The coefficient of determination (R-square) of the q profiles is 99.6%. Compared with the target value, the model results show the potential of the neural network model for practical use in plasma modeling and real-time control of tokamak operations.The main contributions of this work include (i) reconstructing high-resolution (129×129) poloidal f lux distributions directly from edge electromagnetic diagnostics, (ii) introducing a plasma-focused weighted loss function to improve reconstruction quality in the plasma region of interest, and (iii) demonstrating accurate and fast reconstruction of LCFS, X-point and q-profiles with real-time capability.