Neural network tokamak equilibria with incompressible flows

Abstract

We present several numerical solutions to a generalized Grad–Shafranov equation (GGSE), which governs axisymmetric plasma equilibria with incompressible flows of arbitrary direction, using fully connected, feed-forward, deep neural networks, also known as multi-layer perceptrons. Such artificial neural networks (ANNs) are trained to approximate tokamak-relevant equilibria upon minimizing the GGSE mean squared residual in the plasma volume and the poloidal flux function on the plasma boundary. Solutions for the Solovev and the general linearizing ansatz for the free functions involved in the GGSE are obtained and benchmarked against known analytic solutions. We also construct a nonlinear equilibrium incorporating characteristics relevant to the high confinement mode. In our numerical experiments, it was observed that changing the radial distribution of the training points has a surprisingly small effect on the accuracy of the trained solution. In particular, it is shown that localizing the training points at the plasma edge results in ANN solutions that describe quite accurately the entire magnetic configuration, thus demonstrating the interpolation capabilities of the ANNs.