Uncertainty Quantification

Attaching honest error bars to every Kronos prediction by propagating input, data, and model uncertainty through the workloads.

A number without an error bar is incomplete

No simulation output is exact. A breeder TBR of 1.5 or a burner stability margin means little without knowing how uncertain it is. Uncertainty quantification (UQ) is the L0 discipline of attaching honest error bars to predictions by propagating the uncertainties in inputs, data, and models through to the outputs. It is what keeps Kronos results defensible.

Sources of uncertainty

Uncertainty enters from several places: statistical noise in Monte Carlo, uncertainty in nuclear cross-section data, uncertainty in plasma profiles and boundary conditions, and model-form error in surrogates and reduced models. UQ distinguishes these because they behave differently: statistical noise shrinks with more compute, while data and model uncertainty do not.

• Statistical noise from stochastic sampling
• Nuclear data and cross-section uncertainty
• Input and boundary-condition uncertainty
• Model-form error in surrogates and closures

Propagation methods

Kronos propagates uncertainty by running ensembles: sampling the uncertain inputs, running the workload for each sample, and building a distribution over the output. This is why sweeps and replicas matter beyond exploration, they are also UQ. Where full ensembles are too costly, surrogates and sensitivity analysis estimate how strongly each input drives the output.

Sensitivity analysis is a key product. It ranks which uncertainties matter, telling the design team where better data or tighter control would most reduce output uncertainty. For the breeder, it might show TBR is most sensitive to a particular cross-section; for the burner, that confinement is most sensitive to plug density. This turns UQ into design guidance.

UQ also disciplines the twin. Every surrogate carries a validity domain and error estimate, so the real-time twin knows the confidence of its own predictions and can defer when uncertainty is high. Honest error bars, computed offline, are thus what let the fast layers act responsibly on foundation-derived knowledge for both machines.