Uncertainty Propagation in Code Workflows
Uncertainty propagation carries input uncertainties through a chain of codes to produce honest error bars on the final predictions.
Predictions need error bars
A fusion prediction that chains equilibrium, transport, heating, and edge codes rests on uncertain inputs: measured profiles, atomic data, transport-model coefficients, and geometry. Uncertainty propagation tracks how these input uncertainties combine and grow through the workflow, so the final answer, say a predicted fusion power, comes with a defensible confidence interval rather than a single number stated as fact.
Without this step a simulation result is incomplete. Honest reporting means stating both the value and its uncertainty, and identifying which inputs dominate it.
Methods
The direct approach is Monte Carlo sampling: draw inputs from their distributions, run the full workflow for each sample, and build the output distribution. It is robust but expensive. Cheaper alternatives include polynomial chaos expansions, which fit a surrogate polynomial in the uncertain inputs, and linearized error propagation when the workflow is nearly linear.
Surrogates make it tractable
Because a full workflow can be costly, uncertainty studies often replace expensive modules with surrogate models trained on a modest set of full runs, enabling thousands of samples at low cost. The surrogate's own error must then be included in the budget.
Design relevance
For the Hyperion breeder, uncertainty propagation converts frozen design values such as Q 3.424 and 88.7 MW into ranges reflecting model and input uncertainty, supporting the honest position that these are simulation predictions, not guaranteed hardware performance before first tritium around 2030.
- Carries input uncertainty through the code chain
- Monte Carlo, polynomial chaos, or linearized methods
- Surrogates make sampling affordable
- Turns point predictions into confidence intervals