Cross-Machine Transfer and Domain Shift
Why models trained on one tokamak degrade on another, and how transfer learning helps.
The problem
A model trained on one device rarely works as well on another. Machines differ in size, field, wall material, diagnostics, and operating range, so the data distributions differ, a domain shift. This limits how far any single-machine model, such as a disruption predictor, can travel.
Why it matters especially in fusion
There are few large devices, each produces limited data, and the machines that matter most for the future do not yet exist. Any model useful for a new or future machine must transfer across this gap, or be trained on data that machine cannot yet provide.
Transfer techniques
- Fine-tuning: retrain a source model on limited target-machine data
- Domain adaptation: align feature distributions across machines
- Feature engineering toward machine-independent, normalized quantities
- Multi-machine training to learn what generalizes
Machine-independent features
Models built on dimensionless or normalized physics quantities transfer better than models on raw engineering signals, because the physics is more nearly shared across devices. Choosing such features trades some machine-specific accuracy for portability, often a good bargain when the target machine has little data.
Honest expectations
Transfer reduces but does not eliminate the gap, and it works best when source and target overlap in regime. Extrapolating a model to a machine or regime with no overlap, for example a future reactor, is genuinely unsupported by data and should be stated as such. For design-stage concepts, cross-machine learning informs modeling but does not substitute for the physics-based case.