Ml For Fusion

Out-of-Distribution Detection

Recognizing when a model is being asked about inputs unlike anything it was trained on.

The core risk

A model is reliable only on data resembling its training set. Fusion routinely pushes into new regimes, new machines, new operating points, so knowing when an input is out of distribution (OOD) is as important as the prediction itself. An OOD flag says: do not trust this prediction.

Why standard models fail silently

Kronos motion — fusion

Neural networks often produce confident outputs for inputs far from their training data, giving no warning. This is the dangerous failure mode: the model is wrong and confident at once. Explicit OOD detection is the guard against it.

Detection methods

Distance to training data in a feature space
Density estimates that flag low-probability inputs
Ensemble disagreement, high where data are unfamiliar
Reconstruction error from an autoencoder trained on normal data

Use in the loop

In real-time control or monitoring, an OOD flag can trigger a fallback controller, defer to a conservative action, or alert an operator. In offline analysis, it marks predictions that should not be relied upon. The point is to convert silent extrapolation into an explicit, actionable signal.

Connection to design

When models built on existing-device data are applied to a design in a new regime, such as a compact spherical tokamak or a mirror configuration, much of the operating space may be out of distribution by construction. Honest use acknowledges this: OOD detection quantifies how far a prediction reaches beyond the data, and design decisions weight such predictions accordingly rather than treating them as measurements.