Answers are grounded in this paper and cite their sources.
The prediction, mitigation, and avoidance of disruptions are critical prerequisites for ensuring the safe operation of future fusion reactors. While machine learning-based disruption prediction techniques have achieved high accuracy in recent years, their application to a new device demands strong extrapolation capability across parameter regimes. This capability is essential to maintain robust algorithm performance as the operational space of the device expands, while it is also a main shortcoming for most machine learning techniques. This study evaluates the extrapolation capability of a disruption prediction algorithm using multiple subsets of experimental data from HL-3 tokamak, divided by key plasma parameters, yielding three key findings. Firstly, the area under the receiver-operator characteristic curve (AUC) of a standard deep learning algorithm trained on the subset with low parameters degrades to around 0.750, 0.666, and 0.540 when extrapolated to three progressively distant extrapolation regimes. However, the Predict-First Neural Network (PFNN) maintains an AUC of about 0.873, 0.823 and 0.701. Secondly, the algorithm's accuracy shows a clear contrast in degradation behavior: when extrapolating along electromagnetic parameters like plasma current, performance declines slowly and continuously; when extrapolating along energy-related parameters like stored energy, performance remains initially stable but deteriorates abruptly beyond a certain threshold. Finally, after rigorously accounting for the impact of parameter regime extrapolation and adjusting the algorithm deployment strategy, deep learning proved effective in protecting the device during HL-3's high-current and high-beta commissioning phases. This study provides valuable experience for implementing disruption prediction algorithms on new devices and is expected to offer practical reference for the initial operation of future fusion facilities like ITER.
No takes yet. Share an insight, caveat, or question.
Yang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: