AI Plasma Control

Model Predictive Control for Plasmas

Model predictive control optimizes actuator commands over a short future horizon each cycle, handling constraints and coupling directly.

The idea

Model predictive control (MPC) uses a model to predict how the plasma will respond to a sequence of future actuator commands, then chooses the sequence that best meets the goals while respecting constraints. It applies the first command, then repeats the whole optimization next cycle with fresh measurements. This is called receding-horizon control.

Why it suits plasma control

Kronos motion — control room

Plasma control is full of constraints (actuator limits, stability boundaries) and coupling (one actuator affects many outputs). MPC handles both directly: constraints are part of the optimization, and the model captures the coupling. A classical single-loop controller must approximate these; MPC treats them explicitly.

The computational cost

Solving an optimization every cycle is expensive, which historically kept MPC out of the fastest loops. Advances in fast solvers, and the use of simple reduced models, now let MPC run on slower-to-medium loops such as profile and shape control. The fastest loops, like vertical stabilization, still use lightweight controllers where every microsecond counts.

python

# Receding-horizon step
for k in range(N):
    u_seq = solve_qp(model, x_now, refs, u_limits, x_limits)  # optimize horizon
    apply(u_seq[0])          # use only the first command
    x_now = estimate_state() # remeasure next cycle, then repeat

Constraint handling as safety

Because MPC keeps predicted states inside limits, it naturally keeps the plasma away from boundaries, complementing disruption avoidance. If a command would breach a limit over the horizon, MPC never issues it. This built-in respect for constraints is a major reason it is attractive for fusion.

In the Kronos program

Kronos evaluates model predictive control for the coupled, constrained loops of the Hyperion breeder, such as profile and shape control, using the shared reduced models. The fastest stabilization loops remain lightweight for latency. These designs are exercised in the flight simulator ahead of operation.