Fusion Reactor Burn Control with Radial Basis Neural Networks: Preliminary Results

In previous work [1] a standard feedforward artificial neural network (ANN) with sigmoidal activation functions was used to demonstrate the capabilities of ANN for the stabilization of burn conditions, at nearly ignited conditions, of a thermonuclear reactor operating in the low temperature region. In this region, the nominal operating point of the fusion reactor is inherently unstable. The purpose of this work is to report the results of the stabilization of this operating point using the same nonlinear fusion reactor model used in Ref. [1], however here we use radial basis neural networks (RBNN) [2] instead of the standard ANN with only sigmoidal units. As before the dynamical evolution of the fusion reactor is represented by the time evolution of the electron density, the relative density of alpha particles and the temperature of the plasma; with an energy confinement time taken from the CDA-ITER scaling law.