Tokamak equilibrium solver by neural network algorithm

Summary form only given, as follows. Tokamak equilibrium analyses have been routinely carried out by solving the Grad-Shafranov equation with various conventional numerical techniques. As a new approach for the equilibrium analysis a neural network equilibrium solver has been proposed from the functional approximation capabilities of the artificial neural network. The neural network equilibrium solver does not have to perform finite differences and coordinate transformations, so its manipulation becomes easy and simple. It may also save computing time significantly for smoothly evolving equilibria from the equilibrium information of the previous time step. Moreover, the implicitly embedded interpolation function of the neural network solver can make it applicable to real-time equilibrium analysis with the aid of on-line learning algorithms. With comprehensive structural variance and manipulation, characteristics of the neural network solver - the structural dependency on external source profile variations, the number of reasonable iterations for the stable convergence, and computational speed etc. - will be presented and elucidated in view of real-time equilibrium analyses.