Disruption Prediction in Tokamaks
Classifiers that estimate, from live diagnostics, whether a tokamak plasma is approaching a sudden loss of confinement.
What a disruption is
A disruption is a fast, largely irreversible loss of plasma confinement. Stored thermal and magnetic energy dumps into the wall and structures over milliseconds, and runaway electrons and halo currents can damage components. Predicting disruptions early enough to act is a central operational problem for tokamaks.
The prediction task
Given a stream of diagnostic signals up to the present, a predictor outputs a probability that a disruption will occur within a warning horizon (often tens of milliseconds). This is a time-series classification problem with strong class imbalance: most time slices are non-disruptive.
- Inputs: plasma current, density, radiated power, locked-mode signals, MHD mode amplitudes, plasma shape and internal inductance
- Output: disruptivity score, thresholded to raise an alarm
- Metrics: true-positive rate, false-alarm rate, and warning time distribution
Model families
Early systems used simple thresholds on individual signals. Later work used neural networks, random forests, and gradient-boosted trees on engineered features. Deep sequence models (recurrent networks, temporal convolutions) learn directly from raw multichannel signals. Cross-machine studies show that models trained on one device degrade on another unless retrained or adapted.
Why it is hard
Disruptions arise from many distinct precursor chains: density-limit, impurity accumulation, tearing modes, vertical displacement. A single detector must cover them all. False alarms waste discharges and erode operator trust; missed alarms risk damage. Warning time must exceed the actuator response time to be actionable.
Realistic evaluation uses forward-in-time held-out shots, reports the warning-time histogram rather than a single accuracy number, and separates prediction (this page) from avoidance (see the linked page).