Application of Deep Reinforcement Learning in Beam Offset Calibration of MEBT at C-ADS Injector-II

Abstract

As a “Strategic Priority Research Program”, high-current superconducting proton driver linac is becoming the best equipment for nuclear waste disposal and cancer treatment. It can accelerate protons to high energy for transmutation of nuclear waste and damaging to cancer cells. However, the deflection of the beam in the MEBT section causes beam loss, which results in low quality beam. In fact, it is caused by being lack of a method to dynamically adjust its own calibration strategy based on the beam position information. This paper uses a novel Asynchronous Advantage Actor-critic (A3C) artificial intelligence method based on Deep Reinforcement Learning (DRL) to find an optimal control strategy in a fickle environment to solve problem that the existing traditional physical and numerical methods cannot calibrate. The final calibration offset by our method reduces to a mean beam offset of 0.4510 mm, which proves that the proposed method has a very competitive advantage in linac beam offset calibration.