Proof-of-Concept of a Reinforcement-Learning- Based RT shimming technique for HTS magnets

We report a newly developed room-temperature (RT) shimming method for high-temperature superconducting (HTS) magnets employing a deep Q-network (DQN), a type of reinforcement learning theory. With only one training session, the shimming control system (agent) learns how to improve the spatial field homogeneity of an HTS magnet and quickly implements the actual shimming process even under various magnetic field distribution conditions based on the experience gained during the training. Various RT shimming simulations with the MATLAB reinforcement learning toolbox were conducted to verify the feasibility of the method. An agent was trained in a 5 T HTS magnet of which the initial homogeneity was 25.79 ppm at a diameter of 10 mm of the spherical volume (DSV) and enhanced the homogeneity of the magnet under identical field condition. The trained agent was then subjected to various deteriorated field conditions of 32.97 and 35.48 ppm and successfully improved the homogeneity to the target value within a very short time. Shimming results demonstrate that the homogeneity of the HTS magnets, for which the field conditions fluctuate with time due to the screening-current-induced field (SCF) or instability of the power supply, can be improved quickly and frequently by using the proposed method whenever necessary.