Machine-Learned Disruption Precursors
Classifiers learn the signatures that precede a plasma disruption, providing warning time for mitigation before confinement is abruptly lost.
Disruptions and their cost
A disruption is a sudden loss of plasma confinement that dumps thermal and magnetic energy into the vessel, drives large forces, and can generate runaway electrons. Because the consequences scale with device size and stored energy, reliable warning and mitigation are safety-critical for any large tokamak.
Learning precursors
Disruptions are usually preceded by identifiable signatures: growing magnetohydrodynamic modes, radiation spikes from impurity accumulation, density-limit approach, or locked modes. Supervised classifiers trained on labeled discharges learn to map real-time diagnostic streams to a disruption probability.
- Locked-mode and Mirnov-coil signals for mode growth
- Radiated-power fraction and impurity indicators
- Density relative to the empirical density limit
- Plasma current, internal inductance, and beta trends
Model families
Approaches range from gradient-boosted trees on hand-engineered features to deep recurrent and convolutional networks that ingest raw time series. Feature-based models are interpretable and data-efficient; deep models can capture subtler multi-signal patterns but need more labeled disruptions.
Metrics that matter
Performance is judged not just by accuracy but by warning time and the tradeoff between missed detections and false alarms. A missed disruption is dangerous; frequent false alarms waste discharges. Receiver-operating and warning-time curves guide the operating threshold.
Transfer across machines is a known weakness because disruption pathways depend on device and scenario. Cross-machine and physics-guided methods aim to improve portability to new devices with little native data, which matters for any first-of-a-kind machine. In Kronos design work these are simulation and analysis exercises supporting the breeder line; no hardware exists to disrupt yet, and net-gain is not claimed before first tritium.