A new method for tokamak magnetic control using a neural network as a fast surrogate for an inverse Grad-Shafranov solver has been experimentally validated on the Tokamak à Configuration Variable (TCV). The technique provides a real-time mapping from a desired plasma boundary to the required Poloidal Field (PF) coil currents, enabling precise plasma shaping and adaptive control. According to a preprint released on arXiv, the experiments on TCV demonstrated improved plasma shaping accuracy compared to standard discharge preparation procedures, even without the use of explicit real-time shape feedback. This work, conducted at the Swiss Plasma Center, points toward more flexible and responsive control systems for future fusion devices. Source: arXiv

The control architecture's core is a neural network trained to approximate an optimal control policy for plasma boundary regulation. This network functions as an inverse Grad-Shafranov solver, a computationally demanding task that is a bottleneck for conventional real-time control systems. By providing a near-instantaneous calculation of the necessary PF coil currents for a given target shape, the system sidesteps these computational limits. This fast, model-based feedforward control is paired with classical controllers that enforce essential operational constraints, creating a hybrid system that combines the flexibility of machine learning with the reliability of established methods for tokamak operation. Source: arXiv

The control architecture's core is a neural network trained to approximate an optimal control policy for plasma boundary regulation.

Experimental results from the TCV campaign confirmed the system's effectiveness. The neural network controller achieved superior plasma shaping performance relative to the standard TCV discharge preparation process. A key finding was that a single, trained network performed satisfactorily across a range of different plasma magnetic configurations, including single-null and double-null divertor shapes. This versatility suggests the approach is robust and could reduce the need for extensive, configuration-specific controller tuning, a significant operational overhead in many experimental programs. The research highlights the potential for machine learning to generalize across diverse operational scenarios. Source: arXiv

The system's real-time adaptability was also demonstrated, primarily in simulation but with partial experimental validation. The controller showed it could manage dynamic events, such as adaptive strike point motion and the early termination of a discharge in response to a real-time trigger. This capability is critical for future reactors like ITER, which will need to react swiftly to off-normal events or changing plasma conditions to protect machine components and maintain stable operation. The ability to modify the control target on-the-fly without pre-computation represents a significant step beyond static, pre-programmed discharge scenarios. Source: arXiv