An N-Dimensional Neural Network tool for the real-time optimisation of accelerator parameters

We report on the development of a neural network based optimisation agent for the real-time optimisation of electron beam parameters at the Australian Synchrotron Linac. The system mimics an operator´s decisions to perform an optimisation task when no prior knowledge, other than constraints on the actuators, is available. In this paper we show simulation results in a 3D search space, and give results for the real-time optimisation of the beam transmission¹ and energy spread². The agent is designed to handle any number of parameters and in this study the optimisation was performed by adjusting sets of 2, 4 and 7 focusing magnets at a time. Our experimental results demonstrate very satisfactory performance, with a successful increase in transmission from 65% to 70%, in combination with a decrease in the energy spread from 1.26% to 0.80%.