A new preprint details the successful experimental deployment of a reinforcement-learning-based controller for vertical plasma stability on the EXL-50U spherical tokamak. The work addresses a critical challenge in elongated plasma configurations, where vertical displacement events can terminate discharges and risk component damage. The controller was developed and trained within a high-fidelity simulation environment that integrated physics-based plasma-circuit models with experimental equilibrium data from EXL-50U, facilitating a robust sim-to-real transfer. This validation marks a significant step in applying advanced, data-driven control methods to the complex, real-time demands of magnetic confinement fusion devices. Source: arXiv

Spherical tokamaks achieve high plasma beta and favorable confinement properties through a compact aspect ratio and high plasma elongation. However, this elongation (κ) induces a strong vertical instability that requires a fast-feedback control system. Traditional proportional-integral-derivative (PID) controllers are widely used but can be difficult to tune for optimal performance across a wide range of plasma conditions. The research presented in arXiv:2608.20901v1 benchmarks the reinforcement learning (RL) agent against both the operational PID controller and a model-based linear quadratic regulator (LQR) controller, using the same simulated plant dynamics, actuator constraints, and measurement noise for a direct comparison of control strategies. Source: arXiv

Spherical tokamaks achieve high plasma beta and favorable confinement properties through a compact aspect ratio and high plasma elongation.

Simulation results indicate that the RL controller achieves vertical position tracking accuracy comparable to that of the established PID system. The primary advantage identified in the study was the RL agent's efficiency; it demonstrated consistently lower vertical-stabilization coil effort to achieve the same level of control. This reduction in control effort could translate to lower power requirements and reduced stress on power supplies and magnetic coils in future devices. Such efficiency is a key objective for steady-state or long-pulse tokamak operation, where cumulative hardware stress and power consumption are major engineering concerns. Source: arXiv

To bridge the gap between the simulation and the physical device, the researchers augmented the RL controller with a lightweight integral compensation term. This addition improves robustness against the inevitable mismatches between the physics model and the real-world plant dynamics of the EXL-50U. The successful deployment on the tokamak, as stated in the preprint, validates the entire workflow from simulation-based synthesis to experimental application. This methodology could accelerate the development of high-performance control systems for next-generation devices, including those aiming for net energy gain where plasma stability is paramount. Source: arXiv