Artificial intelligent focusing of a microbeam system based on reinforcement learning

Abstract

Ion microbeam facility is a highly effective tool for precise sample irradiation, ion beam micro-modification, ion beam analysis, and other applications at micron and nanometer scale. However, achieving high-resolution beam spots requires meticulous adjustment of the microslit setting, beam transport and magnetic focusing field, which is even time-consuming for well-trained technicians. Nowadays, most of the beamline instruments and power supplies support remote control and automatic adjustment, which promotes the application of artificial intelligence to microbeam formation. In this work, we simulated the 50 MeV proton microbeam system with Oxford triplet lens configuration using a homemade ion optics package, which can generate data about any number of ions passing through quadrupole magnets. Then, an agent interacted with the system and generated large amounts of data. The data was used to train a deep Q-Network (DQN) model. Ultimately, we used the model to accomplish the intelligent focusing function on the simulated microbeam system. Comparative results show that the error between our model and the classic method is less than 0.3%.