Model Predictive Control for Plasmas
Optimizing actuator commands over a receding horizon while explicitly honoring the plasma's and hardware's hard limits.
Control as repeated optimization
Model predictive control (MPC) uses a model to predict the plasma's response over a short future horizon, then solves an optimization each cycle for the actuator commands that best track the reference while satisfying constraints. It applies the first command, then repeats the whole optimization next cycle with fresh measurements - a receding horizon.
Why it fits fusion
MPC's defining strength is that it handles hard constraints directly: coil current and voltage limits, rate limits, and proximity-to-instability limits enter the optimization as constraints, not afterthoughts. Since plasma control is dominated by staying inside limits with limited, coupled actuators, this is a natural match - the controller reasons about the limits it must respect.
The ingredients
- A prediction model, usually a reduced linear plasma-response model
- A cost function trading tracking error against actuator effort
- Constraints on states and inputs (limits and safety margins)
- A solver fast enough to finish inside the control cycle
- A horizon long enough to anticipate, short enough to compute
The cost of prediction
MPC must solve an optimization every cycle, which is heavier than a fixed control law. For fast loops this demands efficient, bounded-time solvers - typically quadratic programs with guaranteed worst-case iteration counts - so the computation fits the real-time budget. Explicit MPC precomputes the solution as a lookup for the very fastest loops.
Where it is used
MPC is well suited to the slower, constraint-dominated control problems - current-profile control, coordinated shape and current, and scenario-level trajectory following - where its ability to respect limits pays off and the cycle time allows an online solve. It is a natural tool for the actuator-management and scenario layers.