acceptodds
Withdrawn from ICLR 2027

SpectralTF-MAE: A Pre-trained Spectral Foundation Model for Tokamak Fusion Plasma Diagnostics

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Spectroscopic diagnostics provide rich information for characterizing plasma states in tokamak fusion experiments, while existing learning-based approaches are often designed for individual tasks and rely heavily on labeled data. In this work, we propose SpectralTF-MAE, a self-supervised spectral foundation model for tokamak plasma diagnostics. The model is pretrained on paired charge-exchange recombination spectroscopy (CXRS) and X-ray crystal spectroscopy (XCS) measurements through masked spectral reconstruction. To better capture diagnostically important spectral structures, we introduce a peak-aware masking strategy together with a peak-enhanced reconstruction objective, encouraging the model to focus on characteristic spectral regions. A shared Transformer encoder is employed to learn transferable representations across different diagnostic modalities. The pretrained model is further adapted to several downstream tasks, including plasma-parameter prediction, cross-diagnostic spectral generation, and spectral peak analysis. Experiments on EAST diagnostic data demonstrate that the proposed framework achieves competitive performance across different tasks and consistently benefits from self-supervised spectral pretraining. These results indicate the potential of SpectralTF-MAE as a general and transferable representation learning framework for plasma spectroscopic diagnostics.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Nuclear Fusion, Spectral, Large Foundation Model, Tokamak Fusion, Plasma Diagnostics

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_atkWOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The paper focuses on an integration of AI and fusion plasma diagnostics and adopts a self-supervised MAE framework to accomplish diagnostics.

Critical strengths

Overall, the paper is a good shape.

Critical weaknesses
  1. From the perspective of AI domain, the novelty of the paper is limited, by simply combining several existing ingredients in the literature. The paper fails to show the uniqueness of the fusion plasma diagnostics, and its added requirements on the model training. Correspondingly, the paper is difficult to follow for the readers in AI domain.
  2. The final result of the paper is also difficult to understand. The compared baselines are the direct application of pretrained checkpoint, and are inconsistent with the SpectralTF-MAE, which is pretrained on the investigated EAST dataset. Simlar spectral-specific pretrained models, such as SpectralGPT,HyperSIGMA,spatial-spectral MAE, are missing to be compared. So, it is hard to evaluate the performance superiority of the proposed algorithm. On the other hand, one of the claimed contribution is its performance on downstream tasks. But Table 2 and Fig. 1 shows inconsistent results on performance improvement.
  3. The method for partitioning data into training, validation, and test sets shall be clearly specified. The paper states that CXRS and XCS data are paired based on discharge and acquisition time; however, it does not explicitly clarify whether the training, validation, and test sets are strictly partitioned according to shot ID. In the context of tokamak data, spectra from consecutive time points within the same discharge—as well as those from successive discharges sharing similar experimental configurations—are often highly correlated. The authors should provide a clear description of their data partitioning strategy to enhance the credibility of the experimental results.
  4. The paper shall be further polished to meet the high standards of ICLR. There still exists several formatting issues in the manuscript.
Questions for authors
  1. Clarify the uniqueness and novelty of the paper.
  2. Explain the superiority and consistency of the proposed method.
Reviewer_2hwtOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents SpectralTF-MAE, a masked-autoencoder pretraining for CXRS and XCS diagnostics from the EAST tokamak, using a shared Transformer encoder with peak-aware masking and a peak-enhanced reconstruction loss. The learned representations are applied to cross-diagnostic spectral reconstruction, plasma-parameter prediction and spectral peak analysis. Its main contribution is adapting masked pretraining to informative spectral structures and demonstrating benefits across these tasks.

Critical strengths

The paper presents a coherent adaptation of masked autoencoder pre-training to CXRS and XCS spectra, with peak-aware masking and a peak-enhanced loss designed to preserve diagnostically informative spectral structures.

The targeted awareness, which is specific to spectral diagnostics, is clearly advantageous and novel.

The evaluation of the model across three downstream task families provides evidence that the learned representation captures information useful beyond reconstruction alone.

Although the demonstrated scope is narrower than the paper’s title with "foundation-model" framing suggests, the study offers promising evidence for self-supervised representation learning across these two spectroscopic diagnostics.

Critical weaknesses

The authors are overselling the scope of the results with their title and some statements in the paper. The title “A PRE-TRAINED SPECTRAL FOUNDATION MODEL FOR TOKAMAK FUSION PLASMA DIAGNOSTICS” is largely misleading. The work here does not demonstrate the capability of the model to transfer across any other diagnostics on EAST (it is only used and evaluated on CXRS and XCS), nor is it demonstrated on the same diagnostics with other devices. While the results are promising and interesting, they are too narrow for ICLR: results including either a) a more complete coverage of EAST diagnostics and/or b) transfer to CXRS/XCS diagnostics on at least another device, would be required to classify this as a “foundation model for tokamak fusion plasma diagnostics”, and qualify for novelty requirements at a venue like ICLR.

Other comments:

Abstract: “a self-supervised spectral foundation model for tokamak plasma diagnostics” => This is an overstatement and could lead the reader to think the foundation model can be adapted to many other diagnostics across several tokamak devices, which is not at all demonstrated in this work.

Conclusion: “foundation model for tokamak plasma diagnostics” => Same here, this is misleading and not demonstrated by the work.

Section-3.3: “For electron density, two measurements are used: ne1 denotes the solid-source measurement and ne2 is obtained from POINT.” => These two diagnostics need to be properly referenced, and authors should clarify these are independent diagnostics from CXRS and XCS.

Section-4.1: “The complete CXRS input is represented as a 1024 × 12 spectral matrix and XCS is spatially integrated to a 35 × 487 representation” => Authors should clarify that these two dimensions are wavelength bins and spatial location. They should also use the same order as presumably the spatial dimensions are 12 and 35 respectively. ie. use “1024 × 12” and “487 × 35”, or vice-versa.

Tabel-3: What is the “Task-specific Baseline” ? It is not clearly explained or referenced anywhere.

Questions for authors

Please address weaknesses mentioned above.

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

The paper develops an architecture, pre-training strategy, and downstream tasks for fusion spectral data. The model (a small and large version) are compared against several standard deep learning architectures. The results indicate that the model predict 3 of 4 parameters better than other models.

Critical strengths

The lesson for fusion plasma diagnostics appears to be the self-supervised pre-training is a powerful approach for that particular problem. This may be a significant result in that field.

Many models were compared against in Table 2.

Critical weaknesses

As stated in the paper, foundation models for spectral data is well-studied, even if it has not been applied directly to plasma spectroscopy. The new challenges introduced by plasma spectroscopy is not obvious from the abstract/intro/related work. As an application paper, this seems most useful for a plasma spectroscopy audience, rather than a machine learning audience. At the least, the paper should evaluate in comparison to SpectralLLM and MolSpectra (cited in intro); or Chai 2019, Samuell 2021, Lin 2024; 20205; and Lee 2026 (cited in related work) - even if these are task-specific, a comparison of the gains of the foundation model would be informative. Table 3 has just one task-specific baseline, but it is described only as “our task-specific peak-analysis baseline” rather than a published state-of-the-art.

The presentation of the paper was difficult to follow. Without experience in plasma diagnostics, the goal of this work is unclear to me. Clearly, the input data is spectral; but what is the goal of the analysis? What does it mean to characterize a plasma state? Figure 1 shows a picture of a state and how that gets measured via spectroscopy, yet what parameters are defining that state and how will machine learning predict those? The bar graphs in Figure 1 don’t communicate the impact of the results. Many plots (especially Figure 3) are too small to read. These are critical results that I was not able to assess.

Questions for authors

As stated in the paper, using foundation models for spectral analysis has been studied. If it has not been applied to this specific problem, then why not? What is particularly challenging about this problem? What is challenging about existing approaches to fusion plasma diagnostics (what needs to be improved in existing methods)?

T_i is better predicted by ViT of SimMIM than by Spectral-TF-MAE, but quite a bit. Is there an explanation for why the models vary so much on their ability to predict ion temperature?

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

Related papers

Discussion (0)

Sign in to comment.