Revision 2

Accepted · cosmetic change · record

Changes against revision 1:

Title · changed: Revision 1: Can a disruption predictor trained elsewhere protect a new tokamak; Revision 2: Can a disruption predictor trained elsewhere protect a new tokamak?
Plain: Machine-learning programs can warn of an oncoming disruption on the tokamak they were trained on. A new large machine like ITER cannot afford many disruptions to learn from, so the predictor must work almost from the start.
Precise: Data-driven and physics-based predictors reach detection rates near 90 percent or better with warning times of order 0.01 to 0.1 s on single devices. Determine whether a predictor trained on existing devices plus a small number of discharges from a new, larger device meets reactor requirements (e.g. above 95 percent detection, below 5 percent false alarms, warning time longer than the mitigation response time) on that device. An answer is a demonstrated cross-device transfer with these metrics.
Settled by: A blind cross-machine test in which a predictor trained without data from the target device meets the stated metrics on it.
Kind: well-posed
Answer type: yes-no
References: (empty)
Literature status: open
Status note: (empty)
Posed since: (empty)

Decisions