acceptodds
Under review as a conference paper at ICLR 2027

Nuclear Fusion Knowledge Graph Guided Edge-Localized Mode Recognition on EAST

Abstract

Recognizing edge-localized modes (ELMs) and their surrounding plasma states is essential for understanding tokamak operation and supporting plasma monitor- ing. The task requires distinguishing physically different states with similar diag- nostic waveforms, resolving brief events, and learning from strongly imbalanced classes. Human recognition combines diagnostic evidence, physical rules, and experience from previous discharges, providing structured information beyond temporal patterns alone. We propose KGF-ELM, a reliability-aware knowledge fusion framework for point-wise recognition of four ELM-related plasma states on EAST. An ELM knowledge graph abstracts the manual recognition workflow into rules that guide historical case construction from 100 training discharges. The Knowledge Graph Collaborative Fusion Module (KGCFM) brings together physi- cal evidence, prototype activations, relational graph representations, and retrieved cases. It coordinates their predictions using source-quality scores and cross-source agreement, then incorporates knowledge through feature enhancement and gated decision correction. A minority-group safeguard attenuates selected collabora- tive updates, while joint training aligns the knowledge pathways with a replace- able temporal backbone. Experiments on EAST-ELM640 show Macro-F1 gains of 2.4–10.4 percentage points across five backbones, averaging 4.7 points. With CrossLinear, KGF-ELM achieves 91.7% Accuracy, 81.6% Macro-F1, and 98.0% Macro-AUC, leading the compared temporal models on these metrics. Abla- tions show that the complete framework outperforms its evaluated fusion and knowledge-component variants. These results establish structured domain knowl- edge as an effective complement to temporal modeling for ELM-related plasma- state recognition.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Edge-Localized Mode Recognition, Knowledge-Guided Learning, Knowledge Graph, Time-Series Classification, Scientific Machine Learning, Tokamak Plasma

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_rQyhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a domain-knowledge guided fusion framework for point-wise recognition of four ELM-related plasma states on the EAST tokamak. It extracts domain-specific physical evidence, activates prototype rules, encodes knowledge graph, and retrieves historical cases. And then, it integrates these knowledge streams with replaceable temporal backbones via reliability-aware future-level enhancement and gated decision-level correction with a minority-class protection safeguard.

Critical strengths

S1. The model incorporates prior agreement and reliability scoring to mitigate over-correction and rule propagation errors. It also introduces a minority protection mechanism to reduce the risk of rare classes being overwhelmed by collaborative updates.

S2. The paper provides fine-grained ablation studies and sensitivity analyses across backbones and individual knowledge components.

Critical weaknesses

W1. The model combines handcrafted rules, a fixed graph, historical cases, multiple gates, safeguards, and auxiliary losses. But it lacks capacity-matched simpler baselines. MLP/ random-graph controls, rule-perturbation tests, or cross-condition evaluation are needed to justify its complexity, robustness and portability.

W2. The model is evaluated under overlapping windows which may leave the statistical and event-level interpretation incomplete. Predictions associated with overlapping windows are temporally dependent, and the reported pooled metrics treat time points observed in different window contexts as separate evaluation instances.

Questions for authors

Q1. How were the physical thresholds, prototype coefficients, graph-relation weights, and the 100 knowledge-construction discharges selected (e.g., from physical priors, training-set statistics, or validation performance)? Please clarify whether any of these choices were tuned on the validation set and confirm that test results were never consulted.

Q2. Could authors clarify the exact implementations of w/o Graph and w/o Reliability in Table 3 (e.g., which inputs and modules are retained or removed, whether model capacity changes, and whether each variant is independently tuned?).

Q3. How were the safeguard parameters (τ,m) selected? How frequently is the safeguard triggered, and what are its class-wise effects on precisions and recall for both minority and majority classes compared with the safeguard-disabled variant?

Reviewer_s9K1OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a framework KGF-ELM for point-wise estimation of plasma states. It combines knowledge graph based module with sequence learning modules and builds the classification model. Experiments on EAST-ELM640 data are shown, where the proposed framework improves performance in multiple metrics including Macro-F1 with 5 choices of backbone, compared with other baseline methods.

Critical strengths
  • Numerical results show a solid and consistent gain over baseline methods, across multiple choices of the backbone model.
  • The Appendix lists the setup and specifications in great detail, which should be appreciated for enabling reproducibility to the extent possible.
Critical weaknesses
  • This paper primarily demonstrates improved performance for one specific task/field, but the contribution to the general learning community is rather limited. It combines knowledge graph with several standard well-established components from ML, such R-GCN and cross attention, with little new insight into the problem. Also, The authors should try to adapt and apply the proposed framework to other problems, and extract generalizable takeaways from them.
Questions for authors
  • What is the general takeaway from this paper, outside of this specific application domain? For example for a researcher working on a different scientific time-series dataset, do you expect the entire architecture to be able to transfer over? Which components are specific to this problem?
  • The acronym EAST appears in the abstract and many times in the main text but is never defined. Granted I don't belong to that field, but if this paper is intended for the general audience at ICLR, it should properly introduce the relevant background context.
Reviewer_UeKGOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

KGF-ELM augments temporal models for four-state plasma recognition with physical diagnostic evidence, prototype activations, graph-derived representations, historical-case retrieval, and reliability-controlled fusion. A minority-group safeguard acts in windows where the base prediction supports Type-III or Large ELM but coordinated knowledge lowers that support. There it suppresses the collaborative decision correction, while the source-specific and feature-level corrections remain active. Evaluation uses EAST-ELM640 with discharge-level training, validation, and test separation.

Critical strengths
  1. The framework addresses an important application and produces improvements across five temporal backbones. Reported single-run macro-F1 gains range from 2.4 to 10.4 percentage points. This breadth makes the overall knowledge-fusion contribution more credible than an improvement on one backbone alone.
  2. The appendix meaningfully strengthens the experimental account, and these checks address plausible concerns about repeatability and retrieval shortcuts:
    • It reports three-seed full-model results for CrossLinear and Timer-XL.
    • It confirms that historical memory is built from 100 of the 448 training discharges, with validation and test discharges excluded.
    • It shows that single-run gains of 1.6 points over the baselines persist when training retrieval excludes same-discharge cases.
Critical weaknesses
  1. The contribution of graph structure is insufficiently distinguished from simpler knowledge encoding.
    • Design. The graph has fixed topology and learned node representations. Its output is shared across samples; the query only determines how the 30 node representations are pooled. This can be a useful structural prior, but the existing comparisons do not establish that relational message passing beats a matched MLP or learned dictionary using the same physical descriptors, supervision, and parameter budget.
    • Ablation size. Removing graph encoding changes macro-F1 by only 0.4 points for CrossLinear and 0.6 for Timer-XL. These differences are comparable to the full model's seed-to-seed variation (CrossLinear SD 0.4, range 0.7; Appendix G.1).
    • Main-table values versus repeated-run means. The main-table full-model scores are the highest macro-F1 values among the three disclosed runs for both backbones. The three-seed mean gains over the single-run baselines are about 2.1 and 2.3 points rather than 2.4 and 2.5. Reporting the means in the main table would be more representative.
    • These differences require paired repeated comparisons before they can support a strong graph-specific conclusion.
  2. The evaluation establishes contextual offline point-wise recognition more directly than performance on independently reconstructed discharge timelines.
    • Overlapping windows. Overlapping window positions are counted separately, so the same physical time point contributes repeatedly under different contexts. The appendix explicitly discloses this protocol and all methods use it; this is not evidence of leakage or an invalid comparison.
    • Missing repetitions. My principal concern is that the baselines and key ablations are single runs, so uncertainty around the smaller component effects is unresolved. Those effects include reliability-aware control (0.2/0.4 points) and the minority safeguard (Table A8).
    • Per-class effects. Aggregate macro-F1 also leaves separate effects on Type-III and Large ELM unresolved; Appendix G.3 itself notes that class-specific precision and recall are needed for this.
    • Event-level results. Consolidated-timeline and event-level results would strengthen application relevance. However, event detection is an additional evaluation, not a requirement of the stated point-wise task.
Questions for authors
  1. Does graph encoding outperform a matched non-graph encoder supplied with identical knowledge and auxiliary objectives, across paired seeds?
  2. What are the per-class precision and recall, and event-level performance, after overlapping predictions are consolidated into one timeline per discharge?
  3. Do the graph, reliability, and minority-safeguard improvements persist under paired repetitions, with uncertainty estimated at the discharge level? The Table 3 component ablations and the Table A8 safeguard ablation are currently single runs.

Comments on the paper

1 comment
Area_Chair_2mGxOpenReview ↗

Please Respond to the Authors’ Rebuttal

Dear Reviewers,

We are now roughly halfway through the Author–Reviewer Discussion period (which runs through 18 November). The authors have submitted their response and, in some cases, a revised manuscript.

Could you please take a moment to:

  • Read the authors’ rebuttal and any updated version of the paper;
  • Leave a reply to the authors on the points you raised, indicating whether your concerns have been addressed.

Even if your evaluation remains unchanged, a short reply confirming your final position is very helpful, as it allows me to track each reviewer’s stance and write an accurate meta-review.

If you are unable to participate further, please let me know as soon as possible.

Thank you for your time and careful reviewing.

Best regards,

Area Chair, ICLR 2027

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.