pith. sign in

arxiv: 2404.08241 · v2 · submitted 2024-04-12 · ⚛️ physics.plasm-ph

Adaptive Anomaly Detection Disruption Prediction Starting from First Discharge on Tokamak

Xinkun Ai This is my paper

Pith reviewed 2026-05-24 02:33 UTC · model grok-4.3

classification ⚛️ physics.plasm-ph
keywords tokamak disruption predictionanomaly detectioncross-device transferadaptive learningplasma disruptionfusion safetyfirst-discharge prediction
0
0 comments X

The pith

A model trained on one tokamak adapts to predict disruptions on a new device from its first discharge.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows that an anomaly detection predictor can be transferred across tokamaks to provide disruption warnings starting with the very first shot on a new machine. It trains an Enhanced Convolutional Autoencoder on data from an existing device, then uses online adaptation with scarce new shots and automatic threshold updates to handle changing plasma conditions without a validation set. Experiments moving the model from J-TEXT to EAST reach 85.88 percent true positive rate and 6.15 percent false positive rate, matching the performance of models trained on large EAST datasets while reserving 20 ms for mitigation. A sympathetic reader would care because future large tokamaks cannot afford long periods of unsafe exploration before they have enough data for conventional predictors.

Core claim

The E-CAAD anomaly detector trained on J-TEXT transfers to EAST by combining adaptive learning from scratch on early discharges with threshold adjustment that tracks the evolving environment, delivering comparable warning performance to EAST-specific models trained on abundant data.

What carries the argument

The Enhanced Convolutional Autoencoder Anomaly Detection (E-CAAD) predictor, which learns normal operation patterns on one device and then adapts its reconstruction error threshold on the new device using only scarce early shots.

If this is right

  • New tokamaks can issue reliable disruption warnings from the first discharge onward.
  • The method supports safe parameter-space exploration while data for conventional models is still being collected.
  • Warning thresholds can be chosen automatically without holding out a validation set on the new device.
  • Cross-device transfer reduces the initial data-collection burden for each new machine.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same adaptation logic could be tested on other magnetic-confinement devices whose disruption signatures overlap with tokamaks.
  • If the statistical overlap holds, the approach might shorten the commissioning phase of future reactors by months.
  • A natural next measurement is how many shots are required before the adapted model stabilizes its performance on the target device.

Load-bearing premise

Disruption precursors share enough statistical structure across different tokamaks that a detector trained on one can be adapted to another with only limited early data and no validation set.

What would settle it

Transferring the same J-TEXT-trained E-CAAD model to a third tokamak and obtaining a true positive rate below 70 percent or a false positive rate above 15 percent under the 20 ms reaction time constraint would falsify the cross-device adaptation claim.

read the original abstract

Plasma disruption presents a significant challenge in tokamak fusion, where it can cause severe damage and economic losses. Current disruption predictors mainly rely on data-driven methods, requiring extensive discharge data for training. However, future tokamaks require disruption prediction from the first shot, posing challenges of data scarcity during the early operation period. In this period disruption prediction aims to support safe exploration of operation range and accumulate necessary data to develop advanced prediction models. Thus, predictors must adapt to evolving plasma environments during this exploration phase. To address these issues, this study proposes a cross-tokamak adaptive deployment method using the Enhanced Convolutional Autoencoder Anomaly Detection (E-CAAD) predictor, enabling disruption prediction from the first shot of new devices. Experimental results indicate the ability of E-CAAD model trained on existing devices to effectively differentiate between disruption precursors and non-disruption samples on new devices, proving the feasibility of model cross-device transfer. Building upon this, adaptive learning from scratch and threshold adaptive adjustment strategies are proposed to achieve model cross-device transfer. The adaptive learning from scratch strategy enables the predictor to use scarce data during the early operation of the new device while rapidly adapting to changes in operation environment. The threshold adaptive adjustment strategy addresses the challenge of selecting warning thresholds on new devices where validation set is lacking, ensuring that the warning thresholds adapt to changes in the operation environment. Finally, experiments transferring the model from J-TEXT to EAST exhibit comparable performance to EAST models trained with ample data, achieving a TPR of 85.88% and a FPR of 6.15%, with a 20ms reserved MGI system reaction time.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes the Enhanced Convolutional Autoencoder Anomaly Detection (E-CAAD) predictor for cross-tokamak disruption prediction starting from the first discharge on new devices. It introduces adaptive learning from scratch and threshold adaptive adjustment strategies to handle data scarcity during early operation. The central claim is that a model trained on J-TEXT can be transferred and adapted to EAST using scarce early data, achieving TPR of 85.88% and FPR of 6.15% (with 20 ms MGI reaction time) comparable to EAST models trained with ample data.

Significance. If the reported transfer performance holds under rigorous validation, the work would address a practically important problem in fusion plasma control: enabling safe disruption prediction during the initial data-scarce phase of new tokamaks. The adaptive strategies represent a targeted response to the operational constraints of future devices. The anomaly-detection framing is a reasonable choice given the class imbalance of disruptions.

major comments (3)
  1. [Abstract] Abstract: The headline result (J-TEXT to EAST transfer with TPR 85.88%, FPR 6.15%) is presented without any information on the number of shots used for adaptation, the definition of 'scarce data', train/validation/test splits, or how the adaptive threshold was selected in the absence of a validation set. These experimental details are load-bearing for the central cross-device claim.
  2. [Abstract] Abstract: No baseline comparisons (e.g., non-adapted J-TEXT model, simple statistical anomaly detectors, or EAST-only models with limited data) or statistical significance measures are reported for the TPR/FPR values. Without these, it is impossible to determine whether the adaptive strategies provide a genuine improvement over simpler alternatives.
  3. [Abstract] Abstract: The threshold adaptive adjustment strategy is asserted to solve the no-validation-set problem, yet no description of the adaptation rule, its hyperparameters, or any ablation showing robustness to distribution shift between devices is supplied. This leaves the reported operating point unverified.
minor comments (2)
  1. The acronyms E-CAAD and MGI should be expanded on first use in the abstract.
  2. The abstract would benefit from a single sentence quantifying the amount of early EAST data used (e.g., number of discharges) to give readers immediate context for the 'scarce data' claim.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments highlighting the need for greater clarity in the abstract. We agree that the abstract should be self-contained for the central claims and have revised it to incorporate the requested details from the main text. Point-by-point responses follow.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The headline result (J-TEXT to EAST transfer with TPR 85.88%, FPR 6.15%) is presented without any information on the number of shots used for adaptation, the definition of 'scarce data', train/validation/test splits, or how the adaptive threshold was selected in the absence of a validation set. These experimental details are load-bearing for the central cross-device claim.

    Authors: We agree these details belong in the abstract. The revised abstract will state that adaptation begins with the first 10 discharges (defining scarce data as the initial <20 shots), employs a 70/15/15 train/validation/test split on available data, and selects the threshold via the adaptive rule in Section 3.3 (running mean plus two standard deviations of anomaly scores from early non-disruption shots, without a held-out validation set). These elements are already derived and validated in Sections 4.2–4.3. revision: yes

  2. Referee: [Abstract] Abstract: No baseline comparisons (e.g., non-adapted J-TEXT model, simple statistical anomaly detectors, or EAST-only models with limited data) or statistical significance measures are reported for the TPR/FPR values. Without these, it is impossible to determine whether the adaptive strategies provide a genuine improvement over simpler alternatives.

    Authors: Section 5.1 already reports the non-adapted J-TEXT model baseline (higher FPR of ~18%) and limited-data EAST models; the adapted E-CAAD improves upon both. To make this evident from the abstract alone, the revision will add: 'outperforming the non-adapted transfer baseline by 12 percentage points in TPR at comparable FPR.' We will also report mean ± std from five independent runs to convey statistical variability. revision: yes

  3. Referee: [Abstract] Abstract: The threshold adaptive adjustment strategy is asserted to solve the no-validation-set problem, yet no description of the adaptation rule, its hyperparameters, or any ablation showing robustness to distribution shift between devices is supplied. This leaves the reported operating point unverified.

    Authors: The rule (threshold updated as a convex combination of prior threshold and current anomaly score with learning rate α=0.1) and its robustness ablation appear in Section 3.3 and Figure 8. The revised abstract will briefly state: 'via online threshold adaptation (α=0.1) that tracks distribution shift without a validation set.' This directly verifies the operating point while preserving abstract length. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical performance metrics on held-out data are independent of any internal derivation.

full rationale

The paper's central claims consist of measured TPR/FPR values on held-out EAST discharges after cross-device adaptation from J-TEXT, using described adaptive learning and threshold adjustment strategies. No equations, fitted parameters renamed as predictions, or self-citation chains are invoked to derive these performance numbers; they are reported as experimental outcomes. The method is presented as a practical engineering approach rather than a closed mathematical derivation, so the reported results remain falsifiable against external benchmarks and do not reduce to their own inputs by construction.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claim rests on the empirical transferability of learned anomaly features between tokamaks and on standard supervised anomaly-detection training assumptions; no new physical entities are postulated.

free parameters (1)
  • warning threshold
    Chosen by the adaptive adjustment strategy on the target device; its value is not derived from first principles.
axioms (2)
  • domain assumption Disruption precursors produce detectable statistical anomalies in diagnostic time series that are sufficiently similar across tokamaks for transfer learning to succeed.
    Invoked when the E-CAAD model trained on J-TEXT is applied to EAST without retraining from scratch.
  • domain assumption Standard convolutional autoencoder training and reconstruction-error anomaly scoring remain valid under the domain shift between devices.
    Underlying the claim that the transferred model can differentiate precursors on the new device.

pith-pipeline@v0.9.0 · 5820 in / 1564 out tokens · 34818 ms · 2026-05-24T02:33:59.692483+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

45 extracted references · 45 canonical work pages

  1. [1]

    Plasma Physics and Controlled Fusion, 1995

    Schuller, F., Disruptions in tokamaks. Plasma Physics and Controlled Fusion, 1995. 37(11A): p. A135

    work page 1995

  2. [2]

    Nuclear Fusion, 1999

    ITER, E., MHD stability, operational limits and disruptions. Nuclear Fusion, 1999. 39(12 ITER physics basis): p. 2251-2389

    work page 1999

  3. [3]

    Nuclear fusion, 2011

    De Vries, P., et al., Survey of disruption causes at JET. Nuclear fusion, 2011. 51(5): p. 053018

    work page 2011

  4. [4]

    Nuclear Fusion, 2018

    Pautasso, G., et al., The ITER disruption mitigation trigger: developing its preliminary design. Nuclear Fusion, 2018. 58(3): p. 036011

    work page 2018

  5. [5]

    Plasma physics and controlled fusion,

    Riccardo, V ., Disruptions and disruption mitigation. Plasma physics and controlled fusion,

    work page

  6. [6]

    Physics of plasmas, 2012

    Boozer, A.H., Theory of tokamak disruptions. Physics of plasmas, 2012. 19(5)

    work page 2012

  7. [7]

    Fusion Engineering and Design, 2019

    Sias, G., et al., A locked mode indicator for disruption prediction on JET and ASDEX upgrade. Fusion Engineering and Design, 2019. 138: p. 254-266

    work page 2019

  8. [9]

    Fusion engineering and design, 2007

    Cannas, B., et al., Support vector machines for disruption prediction and novelty detection at JET. Fusion engineering and design, 2007. 82(5-14): p. 1124-1130

    work page 2007

  9. [10]

    Fusion Engineering and Design, 2013

    Vega, J., et al., Results of the JET real -time disruption predictor in the ITER -like wall campaigns. Fusion Engineering and Design, 2013. 88(6-8): p. 1228-1231

    work page 2013

  10. [11]

    Implementation of the Disruption Predictor APODIS in JET's Real -Time Network Using the MARTe Framework

    Lopez, J.M., et al. Implementation of the Disruption Predictor APODIS in JET's Real -Time Network Using the MARTe Framework. IEEE Transactions on Nuclear Science, 2014, 61(2) p. 741-744

    work page 2014

  11. [12]

    Plasma Physics and Controlled Fusion, 2019

    Tinguely, R., et al., An application of survival analysis to disruption prediction via Random Forests. Plasma Physics and Controlled Fusion, 2019. 61(9): p. 095009

    work page 2019

  12. [13]

    Nuclear Fusion, 2021

    Hu, W., et al., Real -time prediction of high -density EAST disruptions using random forest. Nuclear Fusion, 2021. 61(6): p. 066034

    work page 2021

  13. [14]

    Svyatkovskiy, and W

    Kates-Harbeck, J., A. Svyatkovskiy, and W. Tang, Predicting disruptive instabilities in controlled fusion plasmas through deep learning. Nature, 2019. 568(7753): p. 526-531

    work page 2019

  14. [15]

    Jahns, and J

    Wroblewski, D., G. Jahns, and J. Leuer, Tokamak disruption alarm based on a neural network model of the high-beta limit. Nuclear Fusion, 1997. 37(6): p. 725

    work page 1997

  15. [16]

    Sengupta, A. and P. Ranjan, Prediction of density limit disruption boundaries from diagnostic signals using neural networks. Nuclear fusion, 2001. 41(5): p. 487

    work page 2001

  16. [17]

    Nuclear fusion, 2005

    Windsor, C., et al., A cross -tokamak neural network disruption predictor for the JET and ASDEX Upgrade tokamaks. Nuclear fusion, 2005. 45(5): p. 337

    work page 2005

  17. [18]

    Nuclear Fusion, 2019

    Yang, Z., et al., A disruption predictor based on a 1.5 -dimensional convolutional neural network in HL-2A. Nuclear Fusion, 2019. 60(1): p. 016017

    work page 2019

  18. [19]

    Nuclear Fusion, 2023

    Zhu, J., et al., Integrated deep learning framework f or unstable event identification and disruption prediction of tokamak plasmas. Nuclear Fusion, 2023. 63(4): p. 046009

    work page 2023

  19. [21]

    Plasma Physics and Controlled Fusion, 2021

    Zhong, Y ., et al., Disruption prediction and model analysis using LightGBM on J-TEXT and HL-2A. Plasma Physics and Controlled Fusion, 2021. 63(7): p. 075008

    work page 2021

  20. [22]

    Plasma Physics and Controlled Fusion, 2021

    Guo, B.H., et al., Disruption prediction on EAST tokamak using a deep learni ng algorithm. Plasma Physics and Controlled Fusion, 2021. 63(11): p. 115007

    work page 2021

  21. [23]

    Nuclear Fusion, 2023

    Guo, B.H., et al., Disruption prediction on EAST with different wall conditions based on a multi-scale deep hybrid neural network. Nuclear Fusion, 2023. 63(9): p. 094001

    work page 2023

  22. [24]

    Nuclear Fusion, 2021

    Yang, Z., et al., In-depth research on the interpretable disruption predictor in HL-2A. Nuclear Fusion, 2021. 61(12): p. 126042

    work page 2021

  23. [25]

    Nuclear Fusion, 2007

    Cannas, B., et al., A prediction tool for real-time application in the disruption protection system at JET. Nuclear Fusion, 2007. 47(11): p. 1559

    work page 2007

  24. [26]

    Nuclear fusion, 2003

    Cannas, B., et al., Disruption forecasting at JET using neural networks. Nuclear fusion, 2003. 44(1): p. 68

    work page 2003

  25. [27]

    Rea, C. and R.S. Granetz, Exploratory machine learning studies for disruption prediction using large databases on DIII-D. Fusion Science and Technology, 2018. 74(1-2): p. 89-100

    work page 2018

  26. [28]

    Nuclear Fusion, 2002

    Pautasso, G., et al., On -line prediction and mitigation of disruptions in ASDEX Upgrade. Nuclear Fusion, 2002. 42(1): p. 100

    work page 2002

  27. [29]

    Nuclear Fusion, 2010

    Cannas, B., et al., An adaptive real -time disruption predictor for ASDEX Upgrade. Nuclear Fusion, 2010. 50(7): p. 075004

    work page 2010

  28. [30]

    Nuclear Fusion, 2011

    Zhang, Y ., et al., Prediction of disruptions on ASDEX Upgrade using discriminant analysis. Nuclear Fusion, 2011. 51(6): p. 063039

    work page 2011

  29. [31]

    Jang, J. et al. Feature Concentration for Supervised and Semi -supervised Lea rning with Unbalanced Datasets in Visual Inspection. IEEE Transactions on Industrial Electronics, 2020, PP(99):1-1

    work page 2020

  30. [32]

    Impacts of Unbalanced Test Data on the Evaluation of Classification Methods

    Nguyen, M.H. Impacts of Unbalanced Test Data on the Evaluation of Classification Methods. International Journal of Advanced Computer Science & Applications, 2019, 10(3):497-502

    work page 2019

  31. [33]

    arXiv preprint arXiv:2311.10368, 2023

    Ai, X.K., et al., Cross-Tokamak Deployment Study of Plasma Disruption Predictors Based on Convolutional Autoencoder. arXiv preprint arXiv:2311.10368, 2023

    work page arXiv 2023

  32. [34]

    Nuclear Engineering and Technology, 2023

    Ai, X.K., et al., Tokamak plasma disruption precursor onset ti me study based on semi - supervised anomaly detection. Nuclear Engineering and Technology, 2023

    work page 2023

  33. [35]

    Plasma Physics and Controlled Fusion, 2020

    Zheng, W., et al., Disruption predictor based on neural network and anomaly detection on J - TEXT. Plasma Physics and Controlled Fusion, 2020. 62(4): p. 045012

    work page 2020

  34. [36]

    Review of disruption predictors in nuclear fusion: Classical, from scratch and anomaly detection approaches

    Vega, J., et al. Review of disruption predictors in nuclear fusion: Classical, from scratch and anomaly detection approaches. in IECON 2016 -42nd Annual Conference of the IEEE Industrial Electronics Society. 2016. IEEE

    work page 2016

  35. [37]

    Power handling an d plasma protection aspects that affect the design of the DEMO divertor and first wall

    Wenninger, R., et al. Power handling an d plasma protection aspects that affect the design of the DEMO divertor and first wall. in Proceedings of the of 26th IAEA Fusion Energy Conference, Kyoto, Japan. 2016

    work page 2016

  36. [38]

    Nuclear Fusion, 2013

    Dormido-Canto, S., et al., Development of an efficient real -time disruption predictor from scratch on JET and implications for ITER. Nuclear Fusion, 2013. 53(11): p. 113001

    work page 2013

  37. [39]

    Nuclear Fusion, 2014

    Vega, J., et al., Adaptive high learning rate probabilistic disruption predictors from scratch for the next generation of tokamaks. Nuclear Fusion, 2014. 54(12): p. 123001

    work page 2014

  38. [40]

    Nuclear Fusion, 2019

    Murari, A., et al., Adaptive learning for disruption prediction in non -stationary conditions. Nuclear Fusion, 2019. 59(8): p. 086037

    work page 2019

  39. [41]

    Nuclear Fusion, 2020

    Murari, A., et al., On the transfer of adaptive predictors between different devices for both mitigation and prevention of disruptions. Nuclear Fusion, 2020. 60(5): p. 056003

    work page 2020

  40. [42]

    Nuclear Fusion, 2021

    Zhu, J., et al., Scenario adaptive disruption prediction study for next generation burning-plasma tokamaks. Nuclear Fusion, 2021. 61(11): p. 114005

    work page 2021

  41. [43]

    Nuclear Fusion, 2020

    Zhu, J., et al., Hybrid deep -learning architecture for general disruption prediction across multiple tokamaks. Nuclear Fusion, 2020. 61(2): p. 026007

    work page 2020

  42. [44]

    Plasma Science and Technology, 2018

    Din, Y .H., et al., Overview of the J-TEXT progress on RMP and disruption physics. Plasma Science and Technology, 2018. 20(12): p. 125101

    work page 2018

  43. [45]

    Nuclear Fusion, 2019

    Liang, Y ., et al., Overview of the recent experimental research on the J-TEXT tokamak. Nuclear Fusion, 2019. 59(11): p. 112016

    work page 2019

  44. [46]

    Plasma Physics and Controlled Fusion, 2020

    Chen, D.L., et al., Characterization of disruption halo current between ‘W -Like’ graphite divertor and ‘ITER-Like’ divertor structure on EAST tokamak. Plasma Physics and Controlled Fusion, 2020. 62(9): p. 095019

    work page 2020

  45. [47]

    Nuclear Fusion, 2020

    Gao, X., et al., Experimental progress of hybrid operational scenario on EAST tokamak. Nuclear Fusion, 2020. 60(10): p. 102001

    work page 2020