Uncertainty Propagation in Fusion Design
Fusion performance predictions inherit uncertainty from physics inputs and models; propagating it turns single figures into honest, conditioned ranges.
Why ranges, not points
A fusion design prediction depends on transport coefficients, atomic and nuclear data, geometry tolerances, and model choices - all uncertain to some degree. Reporting a single performance number ignores that spread and invites overconfidence. Propagating the uncertainty produces a distribution of outcomes, so a prediction becomes a range with a stated confidence and stated conditions.
The propagation pipeline
- Characterize input uncertainties as distributions from data and literature
- Reduce dimensionality by screening the inputs that dominate
- Build a validated surrogate of the expensive multiphysics model
- Propagate via Monte Carlo, quasi-Monte Carlo, or polynomial chaos
- Report output distributions, credible intervals, and sensitivity rankings
Separating what data can fix
Good fusion UQ separates epistemic uncertainty - imprecisely known parameters and model form, which better data and physics can reduce - from aleatoric variability that is inherent. This tells designers where more experiments and better models will tighten predictions and where the spread is irreducible, focusing effort where it pays.
Sensitivity guides the program
Sobol and Morris analyses reveal which physics inputs contribute most of the variance in predicted performance. Those inputs become priorities for measurement and model refinement. Inputs that barely move the output can be fixed at nominal values, simplifying the design study.
Guardrails on claims
Uncertainty propagation constrains what can honestly be claimed. Until a machine is built and measured, performance figures are simulated predictions with error bars, not demonstrated results. Reporting the conditions and the interval alongside any central value is what keeps the claim defensible.
In Kronos practice
For the machines, uncertainty in physics inputs is propagated through validated surrogates to report predicted performance as conditioned ranges rather than bare numbers. Frozen design-point figures for the Hyperion breeder and the burner are stated with their conditions, epistemic and aleatoric contributions are distinguished, and no hardware net-gain claim is made before first-of-a-kind results exist.