Uncertainty Quantification in Twins
A twin's predictions are only useful with honest error bars; uncertainty quantification produces and propagates those bars.
Why every twin output needs a range
A twin exists to inform decisions, and a decision needs to know not just the best estimate but how much to trust it. Uncertainty quantification, UQ, is the systematic production of calibrated uncertainty on every state estimate, forecast, and recommendation the twin makes. A prediction without an error band is not actionable; it is a guess dressed as a fact.
Sources of uncertainty
- Measurement noise in the sensors feeding the twin
- Model error where the physics is approximate or a surrogate is imperfect
- Parameter uncertainty in coefficients that are not precisely known
- Aleatory variability, the genuine randomness in the system itself
Aleatory versus epistemic
UQ distinguishes uncertainty that is irreducible, the aleatory part inherent in the process, from uncertainty that more data or better models could reduce, the epistemic part. The distinction matters because it tells operators whether an uncertain forecast can be sharpened by gathering more information or must simply be managed.
Propagation
Uncertainty must be carried through every step, from sensors through state estimation, forward prediction, and into recommendations, so the final error band reflects all contributing sources. Filters propagate it in their covariances; ensembles propagate it through spread; and surrogates must add their own approximation error. Truncating this chain anywhere produces false confidence.
In the Kronos twins
Given that both machines are pre-construction, UQ carries extra weight: much of the twin's uncertainty today is epistemic, reflecting that the physics is validated against related experiments rather than against the machines themselves. Reporting that honestly, and shrinking it as data arrives after first tritium near 2030, is central to the program's credibility. Frozen design numbers such as the Hyperion Q of 3.424 or the burner's 26.49 tesla plug are stated as design points, and the twin's job is to bound how real operation will differ. See validation and ensemble forecasting.