This paper presents Gym-TORAX, a Python package to define Reinforcement Learning (RL) environments for plasma control in tokamaks. Gym-TORAX instantiates a Gymnasium environment from an action space, a state-observation space, and a reward function that measures plasma characteristics. The environment computes plasma states using the TORAX plasma simulator and the objective is to maximize the expected sum of rewards. This plasma control formalization is compatible with most RL algorithms and libraries to facilitate RL research and applications. In its current version, one environment is readily available, based on an International Thermonuclear Experimental Reactor (ITER) scenario. • Reinforcement learning environment in Gymnasium for cross-compatibility. • Reinforcement learning for tokamak plasma control. • TORAX simulations for dynamic plasma modeling.
Mouchamps et al. (Sun,) studied this question.