Surrogate Retraining Cadence

Twin surrogates are retrained on a governed cadence — scheduled, drift-triggered, and regime-triggered — so acceleration never comes at the cost of stale physics.

Fast surrogates need fresh grounding

The twin's neural operators and reduced-order models trade fidelity for speed so the real-time layers can run. That trade is only safe if the surrogates are re-grounded often enough to track the true machine. Retraining cadence is the policy that decides how often each surrogate is refreshed, balancing compute against the risk of a surrogate drifting away from physics.

Cadence is not one number. Fast-changing subsystems (edge conditions, first-wall state) are refreshed more often than slow ones (bulk neutronics response). Each surrogate declares a cadence policy combining a scheduled floor, a drift trigger, and a regime trigger, whichever fires first. The registry tracks the age of each surrogate against its policy so an overdue surrogate is visible.

Cadence drivers

• Scheduled floor: retrain at least every N campaign blocks
• Drift trigger: retrain on shift or concept drift breach
• Regime trigger: retrain when new operating space is explored
• Fidelity trigger: retrain when twin residuals exceed tolerance
• Compute budget: batch overdue surrogates efficiently on L0

def due_for_retrain(sur, now, monitors):
    return (now - sur.trained_at > sur.max_age
            or monitors.drift(sur.name)
            or monitors.new_regime(sur.name)
            or monitors.fidelity(sur.name) > sur.tol)
# batch all due surrogates into one L0 sweep to amortize compute

Cadence connects the continual-learning loop to compute scheduling: retraining is expensive, so overdue surrogates are batched into efficient L0 sweeps rather than retrained one at a time. Every refreshed surrogate re-enters through the validation gates; no surrogate is swapped into the live twin without clearing them.