Tokamak Magnetic Control

A tokamak confines hot plasma with magnetic fields. Changing a coil voltage can affect the plasma’s position, shape and current together. Correcting one error can therefore worsen another, and a sufficiently unstable motion can end the experiment.

One controller faces seven simulated situations: holding shapes, changing current, moving vertically, recovering from a pressure drop and following a combined sequence. The machine is a MAST-U-like simulation, not a demonstration on a real tokamak. Its challenge is sustained feedback: measure, act and correct again as the plasma changes.

The controller does not directly see the plasma boundary. It receives magnetic measurements, coil currents and the requested shape, then sends twelve coil-voltage commands every half millisecond. Training consists of simulated trajectories under different plasma properties, measurement noise and actuator delays.

Evaluation introduces unseen variations, including drifting sensors and faulty power supplies. A controller must distinguish a real change in the plasma from a misleading measurement. This is a test of robustness within a simulator; success would still leave the challenge of transferring to a physical machine.

A successful episode must finish without a coil-current violation and meet every tracking requirement. After allowing the response to settle, mean boundary, gap and X-point-height errors must each be within 1 cm, and current error within 3%. An X-point is a saddle in the magnetic field that helps determine the plasma’s edge geometry.

For example, errors of 0.8, 0.7, 1.1 and 0.9 cm fail because one exceeds its limit. Averaging them would hide the problem. This explains why survival and success are reported separately: a plasma can remain confined while missing the shape needed for an experiment.

The classical reference estimates the plasma shape from magnetic measurements and coordinates the coils to reduce the difference from the target. Separate feedback loops help control current and vertical motion.

The learning starter can add a bounded correction to a conventional PID controller, which responds to the size and history of an error. Learning a correction offers a way to retain stabilizing behavior while improving difficult responses. Alternatively, the model can begin by imitating example commands.

During training, a training-only evaluator, called a critic, can inspect extra simulator information to judge whether an action helped. The deployed controller cannot use those hidden quantities: it must act from magnetic diagnostics and targets. This separation lets rich simulation feedback guide learning without giving the final policy unrealistic senses.

The deployable classical reference keeps 26 of 34 scored episodes alive, but only 11 meet all tracking requirements: about 76% survival versus 32% success. One additional attempt is excluded because of a solver void. These limited results do not establish a learned controller’s advantage.

Two cases also expose a machine-and-metric limitation. The shaping coils are connected in symmetric upper–lower pairs; vertical motion can change the magnetic topology, causing the X-point measurement to select a distant saddle. Even a controller with privileged information encounters this issue. A failed score here needs a physical and measurement diagnosis before it can be blamed on learning alone.

Memory may help identify a drifting sensor or delayed actuator. A useful comparison would test the same controller with and without history under each disturbance, reporting both survival and tracking. Improving nominal accuracy is less convincing if it makes the plasma more fragile when the measurements become unreliable.