Machine-Learning-based 0D Parameter Control on KSTAR

It is required to control 0D plasma parameters such as β_N, β_p, q₉₅, or l_i to sustain the fusion power and stability at the desired level in the future tokamak reactor. In this work, we describe a new scheme to control β_N using the deep reinforcement learning (RL) technique that manipulates multi-dimensional control knobs such as plasma current and boundary shape parameters [1]. Deep RL algorithm trains an artificial decision-making agent to maximize the reward, through its trials and errors of virtual experiments in a data-driven simulator. As the reward is defined to increase when the achieved β_N is close to the target, the RL agent is trained to provide the optimal scenario manipulation for the given target. We address several KSTAR experimental results conducted with RL-determined controls to achieve various targets of β_N. They show successful control of β_N by manipulating the plasma shape only, and also the achievement of high performance of β_N~3 and H₈₉~2.5. Furthermore, we trained the RL agent to control multiple 0D parameters, β_p, q₉₅, and l_i simultaneously into arbitrarily given target regimes, which can provide guidance of the operation scenario development [2]. This methodology can be a first step for the autonomous tokamak operation using machine learning techniques.