Control of a negative-ion accelerator source using neural networks

Abstract The tuning and control of the negative-ion source for a neutral-particle beam accelerator is a time-consuming task for skilled human operators. The source is an inherently highly nonlinear system to which conventional control methods cannot be satisfactorily applied. We report on a program to apply neural-network techniques to the modeling and control of a negative-ion source at the Los Alamos National Laboratory. Suitable networks have been shown to be capable of modeling nonlinear processes upon presentation of a training set consisting of input/output examples taken from the process in question. This model can then be used to train another network to control the process to produce a desired result. Simple examples of this are Widrow's pole balancer and truck backer-upper [1]. We are using this technique to try to predict the beam characteristics of the source for given changes in control settings.