Attention-aware convolutional neural networks for identification of magnetic islands in the tearing mode on EAST tokamak

Adi Liu; Ang Ti; Bihao Guo; Chenguang Wan; Chu Zhou; EAST team; Feifei Long; Ge Zhuang; Hailing Zhao; Jinlin Xie

REVIEW 4 major objections 6 minor 4 references

Attention-aware convolutional neural networks for identification of magnetic islands in the tearing mode on EAST tokamak

T0 review · 4 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read An attention-aware convolutional neural network can identify tearing-mode magnetic islands from electron-cyclotron-emission signals with 91.96% accuracy, outperforming a plain CNN by 4.77 points in accuracy, 5.5 in recall, and 3.74 in…

desk verdict A legitimate, modest classifier paper with a real shot-level test split and a plausible central number; main issues are transparency (code/data, label independence) and internal metric inconsistencies, not the basic approach. read the letter →

arxiv 2412.13498 v1 pith:IS5RL4DM submitted 2024-12-18 physics.plasm-ph

Feifei Long , Yian Zhao , Yunjiao Zhang , Chenguang Wan , Yinan Zhou , Ziwei Qiang , Kangning Yang , Jiuying Li

show 11 more authors

Tonghui Shi Bihao Guo Yang Zhang Hailing Zhao Ang Ti Adi Liu Chu Zhou Jinlin Xie Zixi Liu Ge Zhuang EAST Team

This is my paper · ORCID

classification physics.plasm-ph

keywords tearingmodemagneticislandelectroncyclotronemissionconvolutionalneuralnetworkattentionmechanismtokamakdeeplearningEAST

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to show that an attention-aware convolutional network can reliably identify tearing-mode magnetic islands from electron-cyclotron-emission (ECE) temperature signals on a tokamak. The authors build a classifier from 68 EAST discharges and report 91.96% accuracy on a held-out set of discharges, with the attention module improving accuracy by 4.77 percentage points, recall by 5.5 points, and F1 by 3.74 points over a plain CNN. If the result holds, it would give fusion control systems a fast, automated way to spot islands as they form, a key step toward preventing disruptions in future reactors. The main caveat is that the same ECE data both feeds the model and helped create the labels, so the reported accuracy depends on those labels being correct and independent.

What carries the argument

The central mechanism is a dot-product attention module attached to the CNN's final feature layer. In this module, a learned query multiplies a key matrix, and a softmax produces attention weights used to weight and sum the value representations at each time point; mathematically it computes $\text{Attention}(Q,K,V)=\operatorname{softmax}(QK^{T})V$. The attention focuses the classifier on the ECE time samples where the flattening and phase reversal of electron-temperature signals mark a rotating magnetic island, which is why the authors say the model gains accuracy and recall over a plain CNN.

What would settle it

Re-label a held-out set of the same 68 discharges using only Mirnov-probe and soft-X-ray signals, with no ECE input, then compare the AM-CNN predictions to those labels; if accuracy on the ECE-free labels is much lower than 91.96%, the original accuracy was inflated by ECE-based label leakage.

Watch

Extended reading notes

Core claim

The central claim is that a CNN augmented with a dot-product attention module (AM-CNN) can detect the presence of tearing-mode magnetic islands from 11 ECE channels spanning the plasma core to edge, reaching 91.96% classification accuracy on a test set of 16 EAST discharges. The improvement over a CNN without attention is quantified as +4.77% accuracy, +5.5% recall, and +3.74% F1-score; at a 0.2170 threshold the model also achieves a true-positive rate of 92.99% and a false-positive rate of 8.9%. The authors attribute the gain to the attention mechanism weighting the time points where the rotating island creates phase-reversed temperature fluctuations across channels, and they note that detection is most reliable when the island is mature and less reliable during seed-island onset.

Load-bearing premise

The ground-truth labels for tearing-mode presence are accurate and independent of the ECE features the model sees, since ECE signals were part of the labeling process and any label errors or leakage could inflate the reported 91.96% accuracy.

Editorial extensions

If this is right

An AM-CNN trained on 11 ECE channels can classify tearing-mode presence in held-out EAST discharges with 91.96% accuracy, supporting real-time island identification.
Adding the dot-product attention module improves accuracy by 4.77 percentage points, recall by 5.5 points, and F1 by 3.74 points over a plain CNN on the same train/test split.
At a decision threshold of 0.2170, the model reaches a 92.99% true-positive rate and an 8.9% false-positive rate on the test set.
The model is most reliable once the magnetic island is mature and the temperature reversal surface is distinct; early-stage and seed-island detection remains less stable.
The authors propose extending the work to locked-mode detection, island width and position estimation, and additional diagnostics such as magnetic probes and soft X-ray systems.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Because ECE feeds both the model and the manual labels, the reported test accuracy may overstate real-world performance; a label source independent of ECE would give a more honest estimate.
The model's weaker performance during island onset suggests that labeling early-stage, small islands as positive is difficult; retraining with finer onset labels could improve seed-island detection.
The architecture is general enough that the same attention-plus-CNN setup could likely be applied to locked-mode detection or to other diagnostics, though the paper does not demonstrate this.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Short version: this is a legitimate, modest incremental result — an attention-augmented CNN for classifying tearing-mode vs non-tearing-mode windows from EAST ECE data. The headline number (91.96% accuracy on a held-out shot-level test set) is real in the sense that it is reproduced by the confusion matrix in Figure 6. The paper is above the median for this genre because the split is by discharge, not by random sample, and the baseline comparison uses the same train and test sets. It is not a breakthrough; the components are standard and the novelty is the application, not the architecture.

What it does well: shot-level split, explicit class balance (30.94% negative / 69.06% positive), denoising and time-slice details, and an honest acknowledgment that early-stage / small-island detection is difficult. The attention mechanism is applied at a sensible place (after the fully connected layer) and the authors give a physical rationale for why attention should help — focusing on phase-inversion and temperature-flattening time points.

The soft spots are real but not fatal. First, internal inconsistencies: the confusion matrix implies TPR=91.94%, FPR=7.98%, F1≈0.9405, while the text reports TPR 92.99%, FPR 8.9%, F1 0.9442, and accuracy 92.41% elsewhere. These need reconciliation. Second, no code or data is released, and there are no multiple-seed runs or confidence intervals, so we cannot tell how stable the 4.77% attention gain is. Third, and most important, the label construction is under-specified. Section 2.1 says labels come from Mirnov, ECE, and SXR and are set by correlating ECE with Mirnov signals. Since the model's input is exclusively ECE, the ground truth is not fully independent of the feature space. This is a classic circularity concern, but it is not damning: if Mirnov and SXR provided the primary label and ECE was only used to confirm timing, then the task is legitimate. The paper needs to say which diagnostic anchored the label and ideally provide a blinded relabeling check or release enough data for an external audit.

One more thing: the test set (shots 61042–86559) is earlier in time than the training set (91364–103486). That is a harder and more honest generalization test, but the authors should state it explicitly; as written it looks like an arbitrary split.

Who is this for? Plasma-control and ML-for-tokamak researchers who want a baseline classifier for ECE-based island detection. A serious referee should engage with it. My recommendation: send to peer review, and require the consistency fixes, a label-independence statement, and ideally code/data release. The core result is plausible; it just needs transparency.

Referee Report

4 major / 6 minor

Summary. The paper presents an attention-aware convolutional neural network (AM-CNN) for binary classification of the presence of tearing-mode magnetic islands using 11 channels of electron cyclotron emission (ECE) from the EAST tokamak. A dataset of 68 discharges between 2016 and 2021 is split by shot into training, validation, and test sets. The authors report a test accuracy of 91.96%, with improvements over a plain CNN in accuracy, recall, and F1. The paper also shows a time-evolution analysis of one test discharge and connects model behavior to the physics of island formation.

Significance. If the central claim holds, the work is a relevant step toward real-time tearing-mode identification on EAST and could inform future ITER control. The paper has notable strengths: a shot-level test split that prevents the most obvious temporal leakage, a physically motivated attention mechanism, and a direct comparison with a plain CNN. The main weaknesses are the internal inconsistency of the reported metrics, the partial dependence of the labels on the same ECE diagnostic used as input, and the absence of uncertainty estimates. These issues currently prevent full confidence in the reported accuracy and in the claimed attention gain.

major comments (4)

[Abstract; Section 3.1; Figure 6] The reported performance numbers are internally inconsistent. The abstract states a classification accuracy of 91.96%; Section 3.1 states an accuracy of 92.41% and F1-score of 0.9442, and also reports TPR 92.99% and FPR 8.9% at threshold 0.2170. The confusion matrix in Figure 6(a) gives TN=900, FP=78, FN=176, TP=2007, which implies accuracy = (900+2007)/3161 = 91.96%, TPR = 2007/2183 = 91.94%, FPR = 78/978 = 7.98%, and F1 ≈ 0.940. These values do not match the text in Section 3.1. The manuscript needs a single consistent set of evaluation numbers, with the operating threshold and class prior specified, so that the central accuracy claim is unambiguous.
[Section 2.1] The label construction creates a potential circularity. The labels are described as obtained from Mirnov probes, ECE, and SXR, and as determined by correlation analysis between ECE signals and magnetic probe signals (Section 2.1). The classifier input is exclusively the same 11 ECE channels (Section 2.2.1). If the labeling rule relies on the same ECE features (phase inversions, temperature flattening) that the model consumes, the reported 91.96% accuracy may partly measure the model's ability to reproduce the labeling heuristic rather than its ability to detect tearing modes under an independent ground truth. This is a load-bearing premise for the central claim. A blinded relabeling test using only non-ECE diagnostics (e.g., Mirnov and SXR), or at minimum a quantitative agreement analysis between ECE-derived and non-ECE-derived labels, is needed.
[Section 3.1; Figure 5] The comparison between AM-CNN and CNN reports point estimates only, with no repeated runs or confidence intervals. The text in Section 3.2 saying 'The AM-CNN model is trained 100 times' appears to refer to 100 training epochs, not to 100 independent training runs. Without multiple seeds or bootstrap confidence intervals, the claimed improvements (accuracy +4.77%, recall +5.5%, F1 +3.74%) are not shown to be statistically significant, and the reader cannot assess the stability of the 91.96% test accuracy.
[Section 2.2.2; Equation (1)] The description of the attention mechanism is not sufficiently detailed to reproduce the method. Equation (1) is the standard dot-product attention formula, but the text states that 'a 2×1 matrix is set as keys' and that the weights are dot-multiplied with the input, without defining the dimensions of Q, K, V, how the 26×11×20002 input is mapped to queries/keys/values, or where exactly the attention is inserted relative to the convolutional and pooling layers. Since the attention module is the central novelty and the claimed gain over plain CNN rests on it, the architecture description needs to be precise enough for an independent implementation.

minor comments (6)

[Abstract; Section 2.1] The dataset size is described inconsistently: the abstract says '2.5 × 10^9 data', Section 1 says '2.5 × 10^9 samples', and Section 2.1 says '11362×220022 samples'. Since 11362 × 220022 ≈ 2.5 × 10^9, the latter is the total number of points across all samples; the terminology should be made uniform and the number of independent samples (11362) should be stated clearly.
[Section 2.1; Figure 1] The chronological ordering of the data split should be stated explicitly: the 16 test shots (61042–86559) are earlier than the training shots (91364–103486). If this is an intentional chronological split, the authors should say so and discuss the resulting domain shift; if not, the split order should be justified.
[Section 2.1] The denoising process applied to the ECE signals is mentioned but not described. A brief description of the denoising method is needed for reproducibility.
[Equation (1)] The attention formula in Equation (1) omits the 1/sqrt(d_k) scaling that is standard in the dot-product attention of the cited Transformer paper [28]. The authors should state whether the scaling is used.
[Section 3.1] The definition of precision in the text ('Precision measures the proportion of instances in which the model predicts positive samples to be positive') is incorrect; precision is the proportion of predicted positives that are true positives. The phrasing should be corrected.
[Figure 2(c)] The percentile histogram in Figure 2(c) is not defined; the text should state whether the percentiles are computed over shots, over time samples, or over all data points.

Circularity Check

1 steps flagged · score 3.0 of 10

Accuracy is measured against labels partly built from the same ECE signals the model uses as input, making the headline benchmark partially self-referential; the attention-module comparison retains some independent content.

self definitional [Section 2.1 (Dataset split and Experimental Setup), with model input described in Section 2.2.1 and Figure 3]
"The labels are obtained from experimental diagnostic systems such as Mirnov probes, ECE systems and SXR diagnostic. ... The label presence of tearing modes is determined by performing a correlation analysis between the ECE signals and the magnetic probe signals."

The classifier consumes exactly the 11 ECE channels, and the ground-truth label is determined by performing a correlation analysis between the ECE signals and the magnetic probe signals. ECE is therefore present on both sides of the evaluation: the label encodes an ECE-based heuristic, and the classifier is scored on reproducing that label from the same ECE data. A network that learns the phase-inversion or flattening features used in labeling will achieve high accuracy by construction, so the 91.96% headline number does not establish detection against an independent tearing-mode ground truth.

full rationale

The paper's central result is an empirical classification accuracy on held-out discharges, so it is not a derivation that reduces to its inputs in the usual mathematical sense. The main circularity concern is the label construction: Section 2.1 states that labels are obtained from Mirnov probes, ECE, and SXR, and that presence is determined by correlation analysis between ECE and magnetic probe signals. Because the model input is exclusively ECE channels, the evaluation is partly self-referential and the headline accuracy may measure agreement with an ECE-based labeling heuristic rather than detection against an independent ground truth. This is a real but partial issue: the labels also involve Mirnov probes, SXR, and manual curation, so the model's 4.77% accuracy gain from the attention module is not vacuous and could still reflect genuine feature extraction. The remaining self-citations ([21], [24], [25]) are background references for prior EAST MHD work and ECE calibration, and they are not load-bearing for the model's architecture or performance claim. There is no fitted parameter renamed as a prediction, no uniqueness theorem imported from the authors, and no ansatz smuggled in via citation. The score of 3 reflects a mild but genuine self-reference in the benchmark, not a complete collapse of the derivation into its inputs.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical entities. Its free parameters are model hyperparameters and a post hoc ROC threshold. The central assumptions are domain assumptions about ECE calibration, label quality, channel coverage, and time-window informativeness. No new force, particle, dimension, or conserved quantity is postulated.

free parameters (2)

ROC operating threshold = 0.2170
Reported TPR/FPR are quoted at this threshold, which is selected from the test-set ROC curve rather than fixed before evaluation.
Network hyperparameters = lr=0.0005, batch=26, early-stopping patience=20, filters=[16,64,128,256], kernels=[16,8,4,2], hidden size=2
Chosen by hand with adaptive scheduling and early stopping; no sensitivity analysis or multiple-seed statistics are reported.

assumptions (4)

domain assumption ECE measured intensity is proportional to T_e^3 for the 11 channels at the relevant frequencies
Used in the Introduction to argue ECE is sensitive to islands; this is a standard radiometer relation cited as [24].
domain assumption Manual labels plus ECE-Mirnov correlation are correct ground truth for tearing mode presence
Section 2.1 states labels are obtained from Mirnov probes, ECE, and SXR and confirmed by correlation analysis; all performance metrics depend on label quality.
domain assumption The 11 ECE channels cover the island radial region across all shots despite toroidal field variation from 1.6 to 2.2 T
Section 2.1 and Figure 2(c) claim coverage of roughly 1.8 m < R < 2.1 m; this is needed so training windows contain island features.
domain assumption Random time slices of 5 to 20 ms capture the 1-20 kHz tearing mode signature at 1 MHz sampling
Section 2.1 describes the slicing and frequency range; the input representation is informative only if this holds.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Attention-aware convolutional neural networks for identification of magnetic islands in the tearing mode on EAST tokamak." pith.science (2026). https://pith.science/paper/IS5RL4DM

@misc{pith2026241213498,
  author       = {Pith},
  title        = {Pith review of: Attention-aware convolutional neural networks for identification of magnetic islands in the tearing mode on EAST tokamak},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IS5RL4DM}},
  note         = {Machine review of arXiv:2412.13498}
}

read the original abstract

The tearing mode, a large-scale MHD instability in tokamak, typically disrupts the equilibrium magnetic surfaces, leads to the formation of magnetic islands, and reduces core electron temperature and density, thus resulting in significant energy losses and may even cause discharge termination. This process is unacceptable for ITER. Therefore, the accurate identification of a magnetic island in real time is crucial for the effective control of the tearing mode in ITER in the future. In this study, based on the characteristics induced by tearing modes, an attention-aware convolutional neural network (AM-CNN) is proposed to identify the presence of magnetic islands in tearing mode discharge utilizing the data from ECE diagnostics in the EAST tokamak. A total of 11 ECE channels covering the range of core is used in the tearing mode dataset, which includes 2.5*10^9 data collected from 68 shots from 2016 to 2021 years. We split the dataset into training, validation, and test sets (66.5%, 5.7%, and 27.8%), respectively. An attention mechanism is designed to couple with the convolutional neural networks to improve the capability of feature extraction of signals. During the model training process, we utilized adaptive learning rate adjustment and early stopping mechanisms to optimize performance of AM-CNN. The model results show that a classification accuracy of 91.96% is achieved in tearing mode identification. Compared to CNN without AM, the attention-aware convolutional neural networks demonstrate great performance across accuracy, recall metrics, and F1 score. By leveraging the deep learning model, which incorporates a physical understanding of the tearing process to identify tearing mode behaviors, the combination of physical mechanisms and deep learning is emphasized, significantly laying an important foundation for the future intelligent control of tearing mode dynamics.

Figures

Figures reproduced from arXiv: 2412.13498 by the authors.

**Figure 1.** Tearing mode dataset includes 11362×11×20002 data collected from 68 shots from 2016 to 2021 on the EAST tokamak. 16 shots selected from discharges #61042 to #86559. The remaining 52 shots, spanning from discharges #91364 to #103486, are further divided into the training and validation sets. Experimental signals of tearing mode on the EAST tokamak have been analyzed for the [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 2.** (a) A Diagrammatic of island structure in tearing instabilities shown in the EAST tokamak. (b) Time traces of ECE signals at different radial positions on the outside of island (point C) and inside of island (point A), the O point of island (point B). (c) Percentile pile bar chart of radial different positions in the databases from 11 channels of ECE. The major radius of EAST tokamak is 1.85 m. The formation of tear… view at source ↗

**Figure 3.** The architecture of the segmentation algorithm showing the design of the tearing mode classifier in a multi-channel attention-aware CNN. For a convolutional layer (26@16*2501), each sample generates 16 feature maps, each with a length of 2501, and there are a total of 26 samples or time slices. One sample denotes 11*20002 data and each with 20002 data points, structured as a 3D array 26*11*20002. This network consis… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

4 extracted references · 4 canonical work pages

[1]

Ji X Q, Xu Y , Hidalgo C, et al

[1]. Ji X Q, Xu Y , Hidalgo C, et al. On the interplay between neoclassical tearing modes and nonlocal transport in toroidal plasmas[J]. Scientific reports, 2016, 6(1): 32697. [2]. Tobias B, Yu L, Domier C W, et al. Boundary perturbations coupled to core 3/2 tearing modes on the DIII-D tokamak[J]. Plasma Physics and Controlled Fusion, 2013, 55(9): 095006....

work page 2016
[1735]

MHD stability, operational limits and disruptions

[7]. Hender, T. C., et al. "MHD stability, operational limits and disruptions." Nuclear fusion 47.6 (2007): S128. [8]. Van Vugt D C, Huijsmans G T A, Hoelzl M, et al. Kinetic modeling of ELM -induced tungsten transport in a tokamak plasma[J]. Physics of Plasmas, 2019, 26(4). [9]. Liu B N, Hu W H, Huang Y , et al. Real-Time Estimation of Vertical Instabili...

work page 2007
[2007]

Schuller F C

[15]. Schuller F C. Disruptions in tokamaks[J]. Plasma Physics and Controlled Fusion, 1995, 37(11A): A135. [16]. La Haye, R. J. Neoclassical tearing modes and their control. Phys. Plasmas 13, 055501 (2006). [17]. Van Den Brand H, De Baar M R, Cardozo N J L, et al. Evaluating neoclassical tearing mode detection with ECE for control on ITER[J]. Nuclear Fusi...

work page 2006
[2023]

Disruptive neoclassical tearing mode seeding in DIII -D with implications for ITER

[10]. Bardoczi L, Logan N C, Strait E J. Neoclassical tearing mode seeding by nonlinear three- wave interactions in tokamaks[J]. Physical Review Letters, 2021, 127(5): 055002. [11]. De Vries P C, Johnson M F, Alper B, et al. Survey of disruption causes at JET[J]. Nuclear fusion, 2011, 51(5): 053018. [12]. Turco F, Luce T C, Solomon W, et al. The causes of...

work page 2022

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.

[1] [1]

Ji X Q, Xu Y , Hidalgo C, et al

[1]. Ji X Q, Xu Y , Hidalgo C, et al. On the interplay between neoclassical tearing modes and nonlocal transport in toroidal plasmas[J]. Scientific reports, 2016, 6(1): 32697. [2]. Tobias B, Yu L, Domier C W, et al. Boundary perturbations coupled to core 3/2 tearing modes on the DIII-D tokamak[J]. Plasma Physics and Controlled Fusion, 2013, 55(9): 095006....

work page 2016

[2] [1735]

MHD stability, operational limits and disruptions

[7]. Hender, T. C., et al. "MHD stability, operational limits and disruptions." Nuclear fusion 47.6 (2007): S128. [8]. Van Vugt D C, Huijsmans G T A, Hoelzl M, et al. Kinetic modeling of ELM -induced tungsten transport in a tokamak plasma[J]. Physics of Plasmas, 2019, 26(4). [9]. Liu B N, Hu W H, Huang Y , et al. Real-Time Estimation of Vertical Instabili...

work page 2007

[3] [2007]

Schuller F C

[15]. Schuller F C. Disruptions in tokamaks[J]. Plasma Physics and Controlled Fusion, 1995, 37(11A): A135. [16]. La Haye, R. J. Neoclassical tearing modes and their control. Phys. Plasmas 13, 055501 (2006). [17]. Van Den Brand H, De Baar M R, Cardozo N J L, et al. Evaluating neoclassical tearing mode detection with ECE for control on ITER[J]. Nuclear Fusi...

work page 2006

[4] [2023]

Disruptive neoclassical tearing mode seeding in DIII -D with implications for ITER

[10]. Bardoczi L, Logan N C, Strait E J. Neoclassical tearing mode seeding by nonlinear three- wave interactions in tokamaks[J]. Physical Review Letters, 2021, 127(5): 055002. [11]. De Vries P C, Johnson M F, Alper B, et al. Survey of disruption causes at JET[J]. Nuclear fusion, 2011, 51(5): 053018. [12]. Turco F, Luce T C, Solomon W, et al. The causes of...

work page 2022