ABSTRACT
We describe a project in which we developed an automated adaptive controller based on the CNLS artificial neural network and evaluated its applicability for the tuning and control of a small-angle negative ion source on the discharge test stand at Los Alamos. The controller processes information obtained from the beam current waveform to determine beam quality. The controller begins by making a sparse scan of the four-dimensional operating surface. The independent variables of this surface are the anode and cathode temperatures, the hydrogen flow rate, and the arc voltage. The dependent variable is a figure of merit that is composed of terms representing the magnitude of the beam current, the stability of operation, and the quietness of the beam. Once the sparse scan is finished, the neural network formulates a model from which it predicts the best operating point. The controller takes the ion source to that operating point for a reality check. The operating data are compared with the predicted data to determine the validity of the model. As real data are fed in, the model of the operating surface is updated until the neural network model agrees with reality. The controller then uses a gradient ascent to optimize the operation of the ion source. Initial tests of the controller indicate that it is remarkably capable. It has optimized the operation of the ion source on six different occasions bringing the beam to excellent quality and stability.
