MPC to FPGA Command Handoff
L3's model-predictive control produces optimized trajectories; L1 validates and executes them, so the intelligent target is realized within hard timing.
The seam between thinking and acting
Kronos's MPC agents at L3 solve an optimization over a 50–100 ms predictive horizon: given the twin's forecast, what trajectory of coil currents, beam power, fueling, or DEC voltages best meets the objectives while respecting constraints? The answer is a setpoint trajectory, not an instantaneous command. L1 is where that trajectory becomes physical action.
The handoff contract
L3 publishes the trajectory as advisory targets tagged with validity time. L1 accepts them only after validation — range, rate, and cross-consistency against the live state — then interpolates and executes them through the deterministic loops and drivers. If a new trajectory does not arrive in time, L1 continues on the last valid one; if targets go stale, it falls back to safe local control. The intelligent layer can never stall the machine.
Why MPC lives above L1
- MPC solve time is data-dependent — it cannot meet a 10 µs deadline.
- It optimizes over a horizon; L1 acts on the current instant.
- Its constraints are soft objectives; L1's clamps are hard limits.
- It can be improved offline without touching L1's certified logic.
This split lets the optimization be as sophisticated as compute allows — anticipating a breeder shape excursion or a burner plug drift — while the guarantee that the machine stays safe and on-time rests entirely on the deterministic layer that consumes its output.
Bumpless and bounded
Handover is bumpless: L1 blends onto a new trajectory without stepping the actuators, and arbitration ensures any protection action instantly overrides the MPC target. The result is that learning and determinism cooperate: L3 raises quality, L1 guarantees safety and timing.