Instability and Event Classification
Recognizing which MHD mode or plasma event is present from diagnostic signatures.
The task
Plasmas host a zoo of instabilities and events: tearing modes, edge-localized modes, sawteeth, Alfven eigenmodes, and more, each with characteristic frequencies, spatial structure, and diagnostic signatures. Classifying which is present, and when, supports control, avoidance, and physics analysis.
Signatures and inputs
- Magnetic pickup coils showing mode frequency and toroidal or poloidal number
- Soft X-ray and ECE arrays showing spatial structure
- Spectrograms revealing frequency chirping and mode coupling
- Fast cameras for edge and filament events
Model approaches
Classifiers range from feature-based (mode number, frequency, amplitude fed to a tree or network) to end-to-end models over spectrograms or raw arrays. Convolutional networks work well on spectrogram images; sequence models track modes that evolve and couple over time.
Label quality
Ground-truth labels usually come from experts, who may disagree on ambiguous cases and cannot label every slice of every shot. Label noise limits achievable accuracy and must be measured. Semi-supervised methods use abundant unlabeled data alongside a smaller labeled set to help.
Why classification matters downstream
Different instabilities call for different responses: a tearing mode may be suppressed with localized current drive, an edge-localized mode managed by pellets or field perturbation. Correct classification lets an avoidance system pick the right actuator. Misclassification wastes the response or applies the wrong one, so calibrated confidence, not just a label, is what downstream logic should consume.