Description
High-performance plasma operation in tokamaks is a key requirement for realizing fusion power generation. Internal Transport Barriers (ITBs) enable improved energy confinement and have been intensively investigated from both modeling and experimental aspects as a promising regime for advanced tokamak operation. It is known that ITB formation can be illustrated based on heating and current drive schemes, which determine the evolution of current density and temperature profiles. However, how much control parameters are effectively utilized in practical scenario design still leaves room for optimization.
Machine learning, especially reinforcement learning (RL), provides a framework to systematically explore how much the effectiveness of profile control can be maximized through sequential decision making. Unlike conventional scenario design based on heuristic tuning, RL allows autonomous exploration of control strategies under given physical constraints. In this study, we demonstrate an RL-based investigation of optimized ITB formation scenarios in tokamak plasmas using the integrated transport simulation code TASK.
The RL agent interacts with the transport simulation and learns control trajectories that modify current density and temperature profile evolution, leading to enhanced ITB formation. Through this approach, the capability of RL to extract non-trivial control strategies is illustrated, suggesting its potential as a complementary tool for systematic and flexible scenario optimization in future tokamak research.