Model-Predictive Control: The Optimization
Kronos's MPC agents plan actuation by solving a constrained optimization over the twin's predicted trajectory at every control cycle.
The receding-horizon problem
Model-predictive control chooses actuation by looking ahead. At each cycle it uses the twin's predictive model to forecast how the plant will respond to a candidate sequence of actuator moves over a horizon N, scores that forecast against objectives and constraints, applies only the first move, then re-plans next cycle. The core optimization is:
# MPC optimal control problem, solved every control cycle
minimize over u_0..u_{N-1}:
J = sum_{k=0}^{N-1} [ (x_k - x_ref)^T Q (x_k - x_ref)
+ u_k^T R u_k
+ du_k^T S du_k ] # move suppression
+ (x_N - x_ref)^T P (x_N - x_ref) # terminal cost
subject to:
x_{k+1} = f_twin(x_k, u_k) # twin dynamics (surrogate)
x_k in X_safe # state constraints (envelope)
u_k in U, du_k in dU # actuator limits and slew rates
x_N in X_terminal # terminal safe set
Q, R, S and P weight state error, actuator effort, move rate and terminal error. The dynamics f_twin come from KRONOS-CTRL's surrogates, so the forecast is physics-grounded but fast enough to solve inside the cycle. The state constraint set X_safe is the certified safe operating envelope, and it is what makes MPC the natural controller for a machine that must never leave that envelope.
Why MPC over simple feedback
MPC handles multi-input multi-output coupling and hard constraints natively. The breeder's shape depends on many coils at once; the burner's confinement couples plug fueling, heating and DEC potential. A PID loop cannot respect a stability-margin constraint the way MPC can encode X_safe directly. And because MPC plans over a horizon, it acts before a constraint is violated rather than reacting after.
MPC is advisory to L1: it produces actuation intents that L1 executes deterministically, and it operates in the millisecond regime above the microsecond hard-real-time path and the hardware failsafe.