Reinforcement Learning for ITER Burn Control*

A reinforcement learning (RL)-based control algorithm is proposed for burn control in ITER tokamak plasmas. The RL agent, trained using the Deep Deterministic Policy Gradient (DDPG) algorithm, regulates ion and electron auxiliary heating alongside deuterium and tritium fueling. The primary objective is to accurately track dynamic target plasma states: ion energy density, electron energy density, total particle density, and tritium fraction. The controller's performance is rigorously assessed within a closed-loop simulation environment, which employs a nonlinear zero-dimensional two-temperature plasma model. This model meticulously separates ion and electron energy dynamics and incorporates key physical processes such as fusion heating, radiation losses, and ohmic heating, all parameterized with ITER-relevant conditions to accurately capture plasma behavior. A sophisticated tracking error-based reward function guides the RL agent's learning, enabling it to effectively follow both static and dynamic reference trajectories. The results provide strong evidence of the DDPG agent's capability to achieve precise burn regulation, further highlighting the significant potential of neural network-based controllers for advanced operational control in next-generation fusion reactors.