Actuator Allocation Across Coupled Systems

Both machines have more actuators than control objectives; Kronos's MPC allocates effort across coils, heating, fueling and DEC to meet goals with margin.

Over-actuation as an asset

The breeder has many poloidal-field and solenoid coils plus heating and fueling; the burner has plug fueling, multiple heating sources and DEC potentials. There are more actuators than independent control objectives, which is an asset: the same shape or confinement target can be met by many actuator combinations, and MPC picks the combination that also respects secondary goals, staying off actuator limits, minimizing strain, keeping reserve for disturbance rejection.

Allocation inside the optimization

Kronos does not use a separate allocation stage; allocation falls out of the MPC objective. The actuator-effort term u^T R u with a per-actuator weight matrix R distributes effort, and constraints keep each actuator within limits with headroom. Weighting a nearly saturated coil more heavily pushes effort to coils with reserve, preserving the ability to respond to the next disturbance.

# per-actuator effort weighting drives allocation
R = diag(r_i)                      # larger r_i -> use actuator i less
r_i = base_i * (1 + saturation_penalty(u_i))   # back off near limits
# result: objectives met while keeping reserve on constrained actuators

Allocation is coupled across physics in the burner. The DEC collector potential serves double duty, energy extraction and ambipolar-boundary setting, so its use is allocated jointly by the plug-density MPC and the Power-Systems control. Kronos solves these with a shared state so the allocation is consistent rather than two controllers fighting over one actuator.

When an actuator degrades or is taken offline, the over-actuation provides graceful reconfiguration: MPC simply re-solves with that actuator constrained out, and the remaining actuators pick up the objective if the envelope still permits. If they cannot, the envelope shrinks and the controller moves to a safe reduced-performance operating point rather than pretending the objective is still reachable.