PID and Modern Control in Fusion
When simple proportional-integral-derivative loops suffice and when model-based multivariable control is needed.
The workhorse PID
A PID controller drives the actuator from three terms of the error: proportional (react to present error), integral (eliminate steady offset), and derivative (anticipate from the rate of change). It needs no explicit model, is easy to tune, and handles many single-input single-output loops well. Much routine plasma control - individual gas valves, simple position loops - uses PID.
Where PID runs out
Plasma control is strongly multivariable and coupled: each coil affects the whole boundary, and shape, position, and current interact. A collection of independent PID loops fights itself when the plant is coupled, because each loop treats the others' actions as disturbances. Here a controller that accounts for the coupling explicitly does far better.
Model-based multivariable control
- State-feedback and optimal (LQR/LQG) control use the plasma-response model to coordinate all actuators at once
- Robust control (H-infinity, mu-synthesis) guarantees stability across model uncertainty
- Model predictive control handles hard actuator limits explicitly
- Gain scheduling adapts the controller as the operating point moves
The trade
Model-based control is more capable but depends on a model that may be wrong, and is harder to understand and certify. PID is limited but transparent and robust to model error. Real stacks mix them: model-based control for the coupled magnetic problem, PID for well-behaved single loops, and simple deterministic logic for the safety-critical reflexes where transparency matters most.
Choosing by layer
The choice tracks the layer. Fast, safety-critical loops favor the simplest law that works, for verifiability. Higher, slower layers can afford the complexity of optimization and prediction. Matching controller sophistication to the layer's speed and criticality is itself a design principle.