Reinforcement learning based RF control system for accelerator mass spectrometry

Accelerator Mass Spectrometry (AMS) is a powerful method for separating rare isotopes and electrostatic type tandem accelerators have been widely used. At SungKyunKwan University, we are developing AMS that can be used in a small space with higher resolution based on cyclotron. In contrast to the cyclotron used in conventional PET or proton therapy, the cyclotron-based AMS is characterized by high turn number and low dee voltage for high resolution. It is designed to accelerate not only ¹⁴C but also ¹³C or ¹²C. The AMS cyclotron RF control model has nonlinear characteristics due to the variable beam loading effect of the acceleration of various particles and injected sample amounts. In this work, we proposed an AMS RF control system based on reinforcement learning. The proposed reinforcement learning finds the target control value in response to the environment through the learning process. We have designed a reinforcement learning based controller with RF system as an environment and verified the reinforcement learning based controller designed through the modelled cavity.