pith. sign in

arxiv: 2606.08462 · v1 · pith:TDPW7FBGnew · submitted 2026-06-07 · ⚛️ physics.plasm-ph

Experiment-free disruption prediction for new devices enabled by synthetic diagnostic data augmentation

Zhiqiang Liu , Fengming Xue , Shiwei Xue , Bihao Guo , Dalong Chen , Wei Zheng , Ping Zhu This is my paper

Pith reviewed 2026-06-27 18:04 UTC · model grok-4.3

classification ⚛️ physics.plasm-ph
keywords disruption predictiontokamaksynthetic data augmentationcross-device transferMHD simulationFourier Domain Adaptationzero-shot predictionplasma diagnostics
0
0 comments X

The pith

Synthetic diagnostic data from MHD simulations of the target device raises zero-shot disruption prediction accuracy from 50% to 57% on J-TEXT.

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

The paper seeks to overcome the lack of experimental data for training disruption predictors on new tokamaks by supplementing existing device data with synthetic signals. It generates these signals by running NIMROD MHD simulations matched to the target device's magnetic geometry and diagnostic setup, then applies Fourier Domain Adaptation to align them with real measurements. A model trained on EAST discharges and tested zero-shot on 1,596 J-TEXT cases improves its accurate early warning rate from 50% to 57%. Sympathetic readers care because future devices such as ITER must operate safely from the first plasma without access to prior disruptive shots for training.

Core claim

By integrating the target device's synthetic diagnostic data with Fourier Domain Adaptation, the zero-shot accurate early warning rate of the model on 1,596 J-TEXT discharges is improved from 50% to 57%, while exhibiting enhanced predictive robustness.

What carries the argument

A synthetic diagnostic framework that processes NIMROD MHD simulation data using the target device's magnetic configuration and diagnostic parameters to produce augmenting signals for cross-device models.

If this is right

  • Disruption prediction models can be trained for new devices before any experimental data from them becomes available.
  • The requirement for large volumes of real disruptive discharges from the target device is reduced.
  • Fourier Domain Adaptation aligns synthetic and experimental signals enough to improve generalization across devices.
  • Predictive robustness increases, supporting safer initial operations on devices where disruptions must be strictly avoided.

Where Pith is reading between the lines

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

  • If the resemblance between synthetic and real signals holds for other simulation codes, the method could be applied beyond NIMROD.
  • Pre-operational training on simulated data might allow virtual testing of prediction thresholds before real tokamak runs begin.
  • The same augmentation strategy could transfer to other plasma monitoring tasks that face cross-device data scarcity.

Load-bearing premise

The synthetic diagnostic signals produced by the NIMROD MHD simulation framework, configured with the target device's magnetic configuration and diagnostic parameters, sufficiently resemble the real experimental diagnostic signals to enable effective data augmentation for disruption prediction.

What would settle it

Removing the synthetic data augmentation or using mismatched simulation parameters and observing whether the early warning rate on the 1,596 J-TEXT discharges falls back to 50% or below.

Figures

Figures reproduced from arXiv: 2606.08462 by Bihao Guo, Dalong Chen, Fengming Xue, Ping Zhu, Shiwei Xue, Wei Zheng, Zhiqiang Liu.

Figure 1
Figure 1. Figure 1: Overview of the proposed zero-shot cross-device transfer framework for [PITH_FULL_IMAGE:figures/full_fig_p019_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Synthetic diagnostic signals as functions of time at toroidal angles: (a) 0 [PITH_FULL_IMAGE:figures/full_fig_p020_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of POLARIS and SXR signals between experimental [PITH_FULL_IMAGE:figures/full_fig_p021_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The dual-branch architecture of the TokaFormer model. Inputs are [PITH_FULL_IMAGE:figures/full_fig_p022_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: ITC training loss as a function of epoch for the image-text contrastive loss pre [PITH_FULL_IMAGE:figures/full_fig_p023_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of latent feature t-SNE distributions before and after applying [PITH_FULL_IMAGE:figures/full_fig_p024_6.png] view at source ↗
read the original abstract

Deep learning based approaches have shown great promise in cross-device disruption prediction for tokamaks, however, the robustness of these models heavily relies on massive amounts of training data. For the upcoming ITER, to ensure the safety of the first plasma and subsequent operations, experimental data should be entirely unavailable initially, and disruptive discharges should be strictly avoided thereafter. This extreme data scarcity inherently conflicts with the data-intensive nature of deep learning algorithms. To address this challenge, we utilize synthetic diagnostic signals from the target device to supplement the experimental data from existing devices for the zero-shot disruption prediction on a new device. The detailed implementation pipeline of this scheme is presented. For experimental validation, a predictive model trained on data from the EAST tokamak is deployed for a zero-shot cross-device experiment on the J-TEXT tokamak. A synthetic diagnostic framework, configured with the diagnostic parameters of the target device, is developed to process NIMROD magnetohydrodynamic (MHD) simulation data based on the target device's magnetic configuration, thereby achieving effective data augmentation. Ultimately, the results demonstrate that by integrating the target device's synthetic diagnostic data with Fourier Domain Adaptation, the zero-shot accurate early warning rate of the model on 1,596 J-TEXT discharges is improved from 50% to 57%, while exhibiting enhanced predictive robustness.

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

2 major / 2 minor

Summary. The paper claims that augmenting EAST experimental data with synthetic diagnostic signals generated by NIMROD MHD simulations configured to J-TEXT geometry and diagnostics, combined with Fourier Domain Adaptation, enables improved zero-shot disruption prediction on a new device; specifically, the accurate early warning rate on 1,596 independent J-TEXT discharges rises from 50% to 57%.

Significance. If the result holds, the work would address a key barrier for ITER-scale devices by showing that physics-based synthetic data can supplement scarce experimental training sets for cross-device deep learning. The validation on a large set of real, independent J-TEXT discharges (rather than synthetic test data) is a concrete strength.

major comments (2)
  1. [Abstract / Results] The headline 7 pp gain (50% → 57%) on the 1,596-shot J-TEXT test set is load-bearing for the central claim, yet the manuscript provides no direct distributional comparison (power-spectrum overlap, MMD, or cross-correlation) between the NIMROD-generated synthetic diagnostics and any held-out real J-TEXT discharges; without this, the improvement cannot be unambiguously attributed to the synthetic augmentation rather than the adaptation step or sampling variance.
  2. [Methods / Results] § on model training and baseline: the abstract reports a concrete numerical improvement but supplies no details on training hyperparameters, baseline implementation, statistical significance testing, or error bars on the 50%/57% figures, leaving the support for the claim preliminary.
minor comments (2)
  1. Clarify the exact definition of 'accurate early warning rate' and the time window used for the metric.
  2. Add a table or figure showing the number of disruptive vs. non-disruptive shots in the training, validation, and test partitions.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive report. The two major comments identify genuine gaps in the current manuscript that limit the strength of the central claim. We address each point below and will incorporate the requested analyses and details in a revised version.

read point-by-point responses

  1. Referee: [Abstract / Results] The headline 7 pp gain (50% → 57%) on the 1,596-shot J-TEXT test set is load-bearing for the central claim, yet the manuscript provides no direct distributional comparison (power-spectrum overlap, MMD, or cross-correlation) between the NIMROD-generated synthetic diagnostics and any held-out real J-TEXT discharges; without this, the improvement cannot be unambiguously attributed to the synthetic augmentation rather than the adaptation step or sampling variance.

    Authors: We agree that the absence of a direct distributional comparison leaves the attribution of the 7 pp gain ambiguous. In the revised manuscript we will add a new subsection (and accompanying figure) that reports (i) power-spectrum overlap between the NIMROD synthetic signals and a held-out set of real J-TEXT discharges, (ii) maximum mean discrepancy (MMD) distances, and (iii) cross-correlation statistics. These metrics will be computed on the same diagnostic channels used for training and will be presented alongside the zero-shot prediction results. revision: yes

  2. Referee: [Methods / Results] § on model training and baseline: the abstract reports a concrete numerical improvement but supplies no details on training hyperparameters, baseline implementation, statistical significance testing, or error bars on the 50%/57% figures, leaving the support for the claim preliminary.

    Authors: We accept that the current manuscript lacks the methodological transparency required to evaluate the reported improvement. The revised version will expand the model-training section to include: the full hyperparameter set (learning rate, batch size, architecture details, early-stopping criteria), a precise description of the baseline implementation, results of statistical significance testing (paired t-test or McNemar test on the 1,596 discharges), and error bars on the 50 % / 57 % figures obtained via bootstrapping or repeated random splits. revision: yes

Circularity Check

0 steps flagged

No circularity; central result measured on independent real discharges with external simulation inputs

full rationale

The derivation chain trains a model on EAST experimental data plus NIMROD MHD simulations (configured with J-TEXT geometry and diagnostics) and reports accuracy on 1,596 held-out real J-TEXT discharges. The reported lift (50% → 57%) is an empirical measurement against external experimental ground truth, not a quantity defined or fitted from the same inputs. No equations reduce a prediction to a fitted parameter by construction, no self-citation is invoked as a uniqueness theorem or load-bearing premise, and the synthetic data source is an independent physics code rather than a re-expression of the target dataset. This satisfies the criteria for a self-contained, non-circular result.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The approach depends on the accuracy of MHD simulations and the effectiveness of domain adaptation, with limited free parameters explicitly mentioned in the abstract.

free parameters (1)
  • Fourier Domain Adaptation parameters
    Parameters for aligning the frequency domains between source and target data are likely tuned or fitted.
axioms (1)
  • domain assumption NIMROD MHD simulations accurately model the relevant plasma dynamics for diagnostic signals
    The paper relies on the fidelity of the simulation to generate useful synthetic data.

pith-pipeline@v0.9.1-grok · 5777 in / 1246 out tokens · 27610 ms · 2026-06-27T18:04:35.730765+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

39 extracted references · 6 canonical work pages · 3 internal anchors

  1. [1]

    De Vries P, Johnson M, Alper B, Buratti P, Hender T, Koslowski H and Riccardo V 2011Nuclear Fusion51053018

  2. [2]

    Strait E, Barr J, Baruzzo M, Berkery J, Buttery R, De Vries P, Eidietis N, Granetz R, Hanson J, Holcomb Cet al.2019Nuclear Fusion59112012

  3. [3]

    Park J H and Pereslavtsev P 2024Applied Sciences14936

  4. [4]

    Yang Z, Liu Y, Zhu X, Chen Z, Xia F, Zhong W, Gao Z, Zhang Y and Liu Y 2023Chinese Physics B32075202

  5. [5]

    Zheng W, Xue F, Chen Z, Chen D, Guo B, Shen C, Ai X, Wang N, Zhang M, Ding Yet al.2023 Communications Physics6181

    2023

  6. [6]

    Seo J, Kim S, Jalalvand A, Conlin R, Rothstein A, Abbate J, Erickson K, Wai J, Shousha R and Kolemen E 2024Nature626746–751

  7. [7]

    Wald Y, Feder A, Greenfeld D and Shalit U 2022 On calibration and out-of-domain generalization (Preprinthttps://arxiv.org/pdf/2102.10395)

    work page arXiv 2022

  8. [8]

    Montes K, Rea C, Granetz R, Tinguely R, Eidietis N, Meneghini O, Chen D, Shen B, Xiao B, Erickson Ket al.2019Nuclear Fusion59096015

  9. [9]

    Zhu J, Rea C, Granetz R, Marmar E, Montes K, Sweeney R, Tinguely R, Chen D, Shen B, Xiao Bet al.2021Nuclear Fusion61114005

  10. [10]

    Shen C, Zheng W, Ding Y, Ai X, Xue F, Zhong Y, Wang N, Gao L, Chen Z, Yang Zet al.2023 Nuclear Fusion63046024

    2023

  11. [11]

    Ai X, Zheng W, Zhang M, Ding Y, Chen D, Chen Z, Guo B, Shen C, Wang N, Yang Zet al.2025 Nuclear Fusion65036011

    2025

  12. [12]

    Zhu J, Rea C, Montes K, Granetz R, Sweeney R and Tinguely R 2021Nuclear Fusion61026007

  13. [13]

    Maljaars E, Felici F, Blanken T, Galperti C, Sauter O, De Baar M, Carpanese F, Goodman T, Kim D, Kim S and et al 2017Nuclear Fusion57126063 Experiment-free disruption prediction by synthetic diagnostic data17

  14. [14]

    Shi L, Valeo E J, Tobias B J, Kramer G J, Hausammann L, Tang W M and Chen M 2016Review of Scientific Instruments8711D303

  15. [15]

    Wang B, Meng X, Wang X, Zhu Z, Ye A, Wang Y, Yang Z, Ni C, Huang G and Wang X 2025 (Preprinthttps://arxiv.org/pdf/2507.05198)

    work page arXiv 2025

  16. [16]

    Hender T, Wesley J, Bialek J, Bondeson A, Boozer A, Buttery R, Garofalo A, Goodman T, Granetz R, Gribov Yet al.2007Nuclear Fusion47S128–S202

  17. [17]

    Liu C, Jardin S C, Qin H, Xiao J, Ferraro N M and Breslau J 2022Computer Physics Communications275108313

  18. [18]

    Hoelzl M, Huijsmans G, Artola F, Nardon E, Becoulet M, Schwarz N, Cathey A, Pamela S, Aleynikova K, Antlitz Fet al.2024Nuclear Fusion64112016

  19. [19]

    Jiao Z, Jiang Z, Li Z, Li K and Yuan J 2022 Implementation and Preliminary Analysis of single/dual-SPI based on NIMROD2022 IEEE 3rd China International Youth Conference on Electrical Engineering (CIYCEE)(Wuhan, China: IEEE) pp 1–7 ISBN 978-1-6654-7732-1

    2022

  20. [20]

    Jiang Z, Yang T, Yuan J, Li C, Ye X, Huang J, Liang Y, Izzo V, Huang M, Tong Ret al.2020 Nuclear Fusion60066004

    2020

  21. [21]

    Gong X, on behalf of EAST Team and Collaborators, The EAST Team:, Song Y, Wan B, Li J, Wan Y, Wu X, Liu Fet al.2024Nuclear Fusion64112013

  22. [22]

    Ding Y, Wang N, Chen Z, Xia D, Yang Z, Chen Z, Zheng W, Yan W, Li D, Zhou S and et al 2024 Nuclear Fusion64112005

    2024

  23. [23]

    Ben-David S, Blitzer J, Crammer K, Kulesza A, Pereira F and Vaughan J W 2010Machine Learning79151–175

  24. [24]

    Ding Y, Zhuang G, Zhang X, Zhang J, Ba W, Wang Z and Pan Y 2009Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment606743–748

  25. [25]

    Yang Y and Soatto S 2020 FDA: Fourier Domain Adaptation for Semantic Segmentation (Preprint https://arxiv.org/pdf/2004.05498)

    work page arXiv 2020

  26. [26]

    Sun B and Saenko K 2016 Deep CORAL: Correlation Alignment for Deep Domain Adaptation Computer Vision – ECCV 2016 Workshopsvol 9915 ed Hua G and J´ egou H (Cham: Springer International Publishing) pp 443–450 ISBN 978-3-319-49408-1 978-3-319-49409-8

    2016

  27. [27]

    Shen C, Zheng W, Guo B, Ding Y, Chen D, Ai X, Xue F, Zhong Y, Wang N, Shen Bet al.2024 Nuclear Fusion64066036

    2024

  28. [28]

    Yang L F, Lan T, Wang S X, Zhang Y, Liu H Q, Jie Y X and EAST Team 2026Plasma Physics and Controlled Fusion68015033

  29. [29]

    Zeng S, Zhu P and Ren H 2023Physics of Plasmas30082504

  30. [30]

    Zeng S, Zhu P, Izzo V, Li H and Jiang Z 2022Nuclear Fusion62026015

  31. [31]

    Huazhong University of Science and Technology 2025 Tokamak plasma synthetic diagnosis platform (ESDP) Chinese Software Copyright, Reg. No. 2025SR2139894

    2025

  32. [32]

    Wang Y, Gao L, Shi P, Xu X, Zhou Y, Yang Q, Yang C, Tao Q, Shen C, Wang Y, Wang L, Chen Z, Xia D, Chen Z, Wang N, Yang Z, Ding Y and Pan Y 2022Plasma Science and Technology 24064001

  33. [33]

    Zeng S, Zhu P, Zhou R and Escande D F 2023Nuclear Fusion63016026

  34. [34]

    2025 Open-adas URLhttps://open.adas.ac.uk/

    2025

  35. [35]

    Masi A, Colombo G, De Stasio M, Breschi M, Caponero M A, Celentano G, Marchetti M, Muzzi L, Polimadei A, Savoldi Let al.2025IEEE Transactions on Applied Superconductivity351–5

  36. [36]

    Huang G, Liu Z, Van Der Maaten L and Weinberger K Q 2017 Densely Connected Convolutional Networks2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (Honolulu, HI: IEEE) pp 2261–2269 ISBN 978-1-5386-0457-1

    2017

  37. [37]

    Vaswani A, Shazeer N, Parmar N, Uszkoreit J, Jones L, Gomez A N, Kaiser L and Polosukhin I 2023 Attention Is All You Need (Preprinthttps://arxiv.org/pdf/1706.03762)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  38. [38]

    Devlin J, Chang M W, Lee K and Toutanova K 2019arXiv preprint arXiv:1810.04805(Preprint https://arxiv.org/pdf/1810.04805) Experiment-free disruption prediction by synthetic diagnostic data18

    work page internal anchor Pith review Pith/arXiv arXiv

  39. [39]

    Radford A, Kim J W, Hallacy C, Ramesh A, Goh G, Agarwal S, Sastry G, Askell A, Mishkin P, Clark Jet al.2021 Learning Transferable Visual Models From Natural Language Supervision (Preprinthttps://arxiv.org/pdf/2103.00020) Experiment-free disruption prediction by synthetic diagnostic data19 Figure 1: Overview of the proposed zero-shot cross-device transfer ...

    work page internal anchor Pith review Pith/arXiv arXiv 2021