Data Augmentation for Fusion
Expanding scarce, imbalanced datasets with transformations and synthetic examples, without inventing physics.
Why augment
Fusion datasets are small, imbalanced, and expensive to grow. Data augmentation creates additional training examples from existing ones, helping models generalize and reducing overfitting, especially for rare classes like specific disruptions or instabilities.
Physically motivated transformations
- Adding realistic sensor noise to signals
- Small time shifts and scaling within physical bounds
- Simulating sensor dropout or calibration offsets
- Mixing or interpolating between similar shots with care
The physics constraint
Unlike images, plasma signals must obey physics. An augmentation that produces a signal the machine could never generate teaches the model false patterns. Every transformation must preserve physical validity, which limits how aggressively augmentation can be applied compared to other ML domains.
Synthetic data from simulation
Simulation is a powerful source of extra training data, especially for rare events that appear seldom in experiments. But simulated data carry the simulator's biases, and a model trained mostly on synthetic data may fail on real data. Mixing real and synthetic data, and validating on real held-out shots, keeps this in check.
Honest limits
Augmentation and synthesis do not create new information beyond what the physics and the original data contain; they help a model use what is there. They cannot substitute for real examples of a regime that has never been observed. Claims about performance must always rest on evaluation against real, held-out data.