Feature Engineering for Plasma Signals
Turning raw diagnostics into physically meaningful inputs that make models more accurate and portable.
Why features matter
The inputs a model sees shape what it can learn. Feeding physically meaningful features, rather than raw signals alone, improves accuracy on small datasets, aids interpretability, and helps models transfer across machines. Feature engineering is where domain knowledge enters a data-driven model.
Common plasma features
- Normalized gradients of temperature and density
- MHD mode amplitudes and mode numbers from magnetic arrays
- Radiated-power fraction and impurity indicators
- Dimensionless parameters: beta, collisionality, normalized current
- Spectral features from Fourier or wavelet transforms of signals
Physics-based normalization
Expressing inputs in dimensionless or machine-normalized form makes them comparable across devices and operating points. A model on such features often transfers better than one on raw engineering signals, because the underlying physics is more nearly shared. This is a deliberate trade of some machine-specific detail for portability.
Learned versus engineered features
Deep networks can learn features from raw signals, sometimes finding structure humans miss. But learned features need more data and are less interpretable. On the small datasets typical in fusion, well-chosen engineered features often win, and a hybrid, engineered features augmented by learned ones, can combine the strengths.
Avoiding leakage
A feature must be computable from information available at prediction time. A feature that secretly encodes the outcome, or that uses future data in a real-time setting, produces impressive validation scores and useless deployed models. Auditing features for leakage is part of building them.