Active Learning and Continual Retraining

Kronos targets its expensive offline solves where the surrogates are weakest, and retrains them as the design evolves and, later, as the plant reveals new data.

Spending compute where it helps

Offline high-fidelity solves (Monte Carlo neutronics, FEM thermomechanics, reference MHD) are expensive, so Kronos does not sample the operating space uniformly. Active learning chooses the next simulations to run where the surrogates are most uncertain or where control most needs accuracy, so each expensive solve maximally improves the twin.

# active-learning acquisition loop (offline, L0)
while budget:
    x_query = argmax_x acquisition(epistemic_uncertainty(x),
                                   control_relevance(x))
    y = high_fidelity_solve(x_query)     # Monte Carlo / FEM / MHD
    dataset.add(x_query, y)
    retrain_surrogates(dataset)

The acquisition function balances epistemic uncertainty (learn where the model is ignorant) against control relevance (prioritize regions the machine actually operates in and the safety-critical edges of the envelope). This concentrates fidelity on the negative-triangularity pedestal for the breeder and the plug-throat/ambipolar region for the burner, exactly where control tolerance is tightest.

Continual retraining

Pre-FOAK the design still changes, so surrogates are retrained in batch at L0 as geometries and parameters update, then re-validated before redeployment. Post-FOAK, plant telemetry becomes the richest data source: divergence between shadow and plant flags where the surrogate is wrong, active learning targets those regions, and retraining folds real data in. This is the loop that keeps the twin faithful as materials activate and the machine ages under neutron fluence.

Retraining is governed, not automatic: a new surrogate version passes the full V&V and calibration gate and is versioned with lineage before it can inform control, so continual learning never silently swaps in an unvalidated model on the safety path.