Model Calibration to Data
Calibration tunes a model's uncertain parameters so its predictions match observations, without overfitting to noise.
Why calibration is needed
Every model has parameters that are not known precisely, such as transport coefficients or material properties. Calibration adjusts these within their plausible ranges so the model reproduces measured data. A calibrated model predicts better, provided the calibration is done honestly and does not simply fit the noise.
How it is done
- Identify which parameters are uncertain and their plausible ranges.
- Compare model predictions to measured data across many conditions.
- Find the parameter values, with uncertainty, that best explain the data.
- Test the calibrated model on data it was not fit to.
Calibration versus fitting
Good calibration respects the physics and reports parameter uncertainty; bad calibration bends parameters until any curve fits, which produces a model that matches the training data and fails everywhere else. The guard is validating on held-out data, part of VVUQ.
Principled inference
When the model can simulate but has no closed-form likelihood, simulation-based inference provides a principled route to calibration that carries full uncertainty, rather than returning a single tuned value.
For Kronos
Before the Hyperion breeder and the burner operate, calibration uses data from other devices and from irradiation tests. After first tritium, targeted around 2030, their own operating data calibrates the models further, turning the flywheel.
Honesty
A calibrated parameter is reported with its uncertainty and the data it was fit to, so its range of validity is clear and it is never treated as a fixed truth.