Twin Fidelity Metrics

Kronos measures how well the twin represents each machine with concrete, per-module fidelity metrics, not a single vague score.

Fidelity is plural

There is no single number for whether a digital twin is good. A twin can predict plasma shape well and thermal loads poorly, or be accurate at flat-top and unreliable during ramps. Kronos defines fidelity per module and per operating regime, with metrics tied to what each consumer needs, so a fidelity claim is always specific and checkable.

• MHD: boundary-position error (mm), delta/kappa error, growth-rate agreement
• Neutronics: flux/spectrum error vs Monte Carlo, TBR error against the 1.1/1.5/1.8 lever
• Thermomechanics: temperature and strain field error vs FEM, divertor heat-flux error
• Power/DEC: potential and efficiency error, grid-sync tracking error
• Shadow: forecast error vs realized state over the 50-100 ms horizon

Regime-resolved

Each metric is reported per regime, commissioning ramp, flat-top, transient, because fidelity that is adequate at steady state can be inadequate during the fast changes where control matters most. Kronos reports the worst-case regime, not the average, since control safety depends on the twin being right when the machine is hardest to model.

Fidelity metrics also drive where effort goes: a module or regime with poor fidelity is where active learning samples and where surrogates are refined. The metrics therefore close a loop, they diagnose weakness, that weakness targets offline compute, and retraining improves the metric, which is re-measured.

Because the machines are pre-FOAK, fidelity today is measured against high-fidelity offline references (Monte Carlo, FEM, reference MHD) and legacy-device data. The metric framework is designed so that when FOAK data arrives ~2030, the same metrics are simply re-evaluated against the real plant, giving a continuous fidelity record across the simulation-to-hardware transition.