Reproduce “Cross-tokamak Disruption Prediction based on Physics-Guided Feature Extraction and domain adaptation” on OpenResearch

About this paper

Chengshuo Shen, Wei Zheng, Bihao Guo, Yonghua Ding, Dalong Chen, Xinkun Ai, Fengming Xue, Yu Zhong and 6 more

This research develops an approach for cross-tokamak disruption prediction by integrating physics-guided feature extraction with supervised domain adaptation, enabling effective prediction in data-scarce target tokamaks. The proposed supervised CORAL (S-CORAL) model achieved an AUC of 0.890 on EAST using only 110 target discharges, closely approaching the performance of a full-data EAST model (AUC 0.936).

Key insights

• Physics-Guided Feature Extraction (PGFE) creates a transferable and interpretable feature space that reduces device-specific variations across different tokamaks.
• Supervised CORAL (S-CORAL) effectively bridges distribution gaps between source and target tokamaks using limited target device data, substantially improving cross-tokamak prediction accuracy.
• Interpretability tools like SHAP can reveal how models learn and highlight device-specific physical differences (e.g., plasma displacement patterns) that persist even after domain adaptation.

Summary by alphaXiv.