pith. sign in

arxiv: 2605.18310 · v1 · pith:HWEEGCSAnew · submitted 2026-05-18 · 🌊 nlin.CD

Comparative blobs and holes dynamics in a tokamak plasma: deep learning analysis of fast imaging data

F Brochard (IJL) , H Aksoy (IJL) , S Chouch\`ene (IJL) , J Cavalier , M Desecure , N Lemoine (IJL) This is my paper

Pith reviewed 2026-05-19 23:39 UTC · model grok-4.3

classification 🌊 nlin.CD
keywords tokamak plasmablobs and holesplasma turbulencefast imagingdeep learning analysismedian subtractionCOMPASS tokamak
0
0 comments X

The pith

Most negative structures in tokamak plasma images are artifacts from sliding-median subtraction rather than real holes.

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

The paper examines turbulent structures in fast passive imaging data from the COMPASS tokamak after tomographic inversion and subtraction of a sliding median image to isolate fluctuations. Positive structures display dynamics consistent with the expected behavior of blobs in plasma turbulence. Negative structures, presumed to be holes, instead show nearly identical dynamics to the positives, which contradicts the theoretical expectation that holes should move in the opposite sense. This mismatch leads to the conclusion that most negative structures arise as artifacts of the median-subtraction preprocessing step. When only supernumerary negative structures are retained, their dynamics align with those anticipated for holes, providing a practical route to isolate genuine hole behavior.

Core claim

After sliding-median subtraction of tomographic images, deep-learning tracking shows positive structures follow blob-like dynamics while the full set of negative structures follows nearly the same dynamics, indicating that the majority of negatives are preprocessing artifacts; retaining only supernumerary negatives yields dynamics consistent with holes.

What carries the argument

Deep-learning comparison of the motion and lifetimes of positive and negative structures identified after sliding-median subtraction from fast imaging frames.

If this is right

  • Positive structures after median subtraction exhibit the radial propagation and lifetimes expected for blobs.
  • The complete population of negative structures behaves too similarly to blobs to be interpreted as physical holes.
  • Filtering to retain only supernumerary negative structures produces dynamics matching theoretical expectations for holes.
  • The filtered subset of negatives offers a workable starting point for systematic study of hole properties in tokamak turbulence.

Where Pith is reading between the lines

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

  • Improved preprocessing that avoids creating motion-mimicking negatives could raise the fraction of usable hole candidates.
  • The same deep-learning pipeline could be applied to data from other devices to test whether the artifact fraction depends on diagnostic or plasma parameters.
  • Reliable isolation of holes would allow direct tests of blob-hole asymmetry in turbulent transport models.

Load-bearing premise

Subtracting a sliding median isolates genuine physical fluctuations and does not systematically create artificial negative features whose motion copies that of blobs.

What would settle it

Repeating the analysis with an independent method of highlighting fluctuations that avoids sliding-median subtraction and checking whether the resulting negative structures then display dynamics opposite to those of the positive structures.

read the original abstract

Abstract This work focuses on the dynamics of the turbulent structures revealed by tomographic inversion of fast passive imaging data acquired on the COMPASS tokamak. To highlight the fluctuations, a sliding median image is subtracted from each image, revealing positive and negative structures. Assuming that the positive structures are blobs and the negative structures are holes, a recently developed deep learning analysis method is used to compare the dynamics of the two types of structures. While the results obtained for the positive structures seem to be in line with the dynamics expected for blobs, contradictory results are obtained for the negative structures, since their dynamics are very similar to those of blobs whereas they should be opposite. Our work suggests that the majority of negative structures resulting from data pre-processing are artefacts produced by the latter. However, a basic approach that only retains supernumerary negative structures shows that the behaviour of the latter is consistent with that expected for holes, opening new perspectives for their investigation.

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 analyzes fast passive imaging data from the COMPASS tokamak after sliding-median subtraction to isolate positive (interpreted as blobs) and negative (interpreted as holes) turbulent structures. A deep-learning tracking method is applied to compare their dynamics. Positive structures exhibit velocities and sizes consistent with expected blob behavior, while negative structures show statistically similar dynamics, leading the authors to conclude that the majority are artifacts induced by the pre-processing step. A supplementary filter retaining only supernumerary negative structures is shown to recover dynamics consistent with holes.

Significance. If the central interpretation is validated, the work would demonstrate that common sliding-median pre-processing can systematically generate artifactual negative features whose tracked properties mimic physical blobs, while also providing a practical filter to isolate candidate holes. This would refine analysis pipelines for fast imaging in fusion plasmas and open quantitative routes to study hole dynamics, which remain less characterized than blobs. The deployment of a pre-trained deep-learning tracker on experimental data is a methodological strength.

major comments (3)
  1. [Abstract and §4] Abstract and §4 (Results): The claim that most negative structures are pre-processing artifacts rests on the observed similarity of their dynamics to those of positive structures. However, no synthetic-data control is reported in which only positive fluctuations are injected into background images before applying the identical sliding-median subtraction; without such a test it remains possible that the filter itself induces negative dips whose subsequent deep-learning outputs (velocity, size) are comparable by construction.
  2. [§3] §3 (Data pre-processing): The assumption that sliding-median subtraction isolates genuine physical fluctuations without systematically generating artifactual negative structures is load-bearing for the artifact conclusion, yet no quantitative metric (e.g., fraction of negative pixels induced on pure-positive synthetic fields, or error bars on the deep-learning outputs) is supplied to bound this effect.
  3. [§5] §5 (Discussion): The supernumerary-negative filter is presented as rescuing a hole interpretation, but the selection criterion is not shown to be independent of the same velocity and size metrics used to compare dynamics; this risks circularity when claiming consistency with expected hole behavior.
minor comments (2)
  1. [Figure 3] Figure 3 and associated text: the deep-learning output distributions would benefit from explicit error bars or confidence intervals derived from the network rather than only mean values.
  2. [§5] Notation: the term 'supernumerary negative structures' is introduced without a precise mathematical definition (e.g., a threshold on intensity or spatial overlap); a short equation or algorithmic step would remove ambiguity.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed comments, which help clarify the strength of our conclusions regarding artifactual negative structures in the COMPASS imaging data. We address each major comment below and commit to revisions that directly respond to the concerns raised.

read point-by-point responses

  1. Referee: [Abstract and §4] The claim that most negative structures are pre-processing artifacts rests on the observed similarity of their dynamics to those of positive structures. However, no synthetic-data control is reported in which only positive fluctuations are injected into background images before applying the identical sliding-median subtraction; without such a test it remains possible that the filter itself induces negative dips whose subsequent deep-learning outputs (velocity, size) are comparable by construction.

    Authors: We agree that a dedicated synthetic control injecting only positive fluctuations prior to sliding-median subtraction would provide direct evidence that the pre-processing step generates artifactual negatives whose tracked properties mimic those of the positives. Our present conclusion rests on the empirical observation that negative structures exhibit statistically indistinguishable velocities and sizes from positives, contrary to the opposite polarity expected for holes. We will add this synthetic test in the revised manuscript, using the same background images and deep-learning tracker to quantify the induced negative features. revision: yes

  2. Referee: [§3] The assumption that sliding-median subtraction isolates genuine physical fluctuations without systematically generating artifactual negative structures is load-bearing for the artifact conclusion, yet no quantitative metric (e.g., fraction of negative pixels induced on pure-positive synthetic fields, or error bars on the deep-learning outputs) is supplied to bound this effect.

    Authors: We acknowledge that the original manuscript lacks explicit quantitative bounds on the artifact generation. We will incorporate metrics such as the fraction of negative pixels produced from synthetic positive-only fields after sliding-median subtraction, together with statistical uncertainties or error bars on the deep-learning-derived velocities and sizes, to better characterize and limit the pre-processing effect. revision: yes

  3. Referee: [§5] The supernumerary-negative filter is presented as rescuing a hole interpretation, but the selection criterion is not shown to be independent of the same velocity and size metrics used to compare dynamics; this risks circularity when claiming consistency with expected hole behavior.

    Authors: The supernumerary-negative filter retains structures whose number or spatial distribution exceeds that attributable to the positive population (i.e., excess negatives without corresponding positive counterparts). This selection is performed prior to and independently of the velocity and size statistics. Nevertheless, to eliminate any appearance of circularity we will add explicit verification that the filter criterion does not incorporate velocity or size information, and we will report additional dynamical quantities (e.g., lifetime, radial propagation) for the filtered negatives to demonstrate consistency with hole expectations beyond the original comparison metrics. revision: partial

Circularity Check

0 steps flagged

No circularity: dynamics comparison driven by experimental images and pre-trained DL tracker

full rationale

The paper subtracts a sliding median from fast imaging data on COMPASS tokamak, then applies a recently developed deep-learning tracker to positive and negative structures. Positive structures yield velocities and sizes consistent with expected blob behavior; negative structures yield similar statistics, leading to the inference that most negatives are pre-processing artifacts. A secondary filter retaining only supernumerary negatives is then shown to recover hole-like behavior. No equation reduces the reported velocities or sizes to a fitted parameter taken from the same dataset, no self-citation supplies a uniqueness theorem that forces the conclusion, and the DL model is described as pre-trained on external data. The central claim therefore rests on direct comparison of independently extracted observables against prior physical expectations rather than on any definitional or self-referential loop.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim depends on standard domain assumptions about the physical identity of positive and negative fluctuations and on the reliability of the deep-learning tracker; no new entities are postulated and no free parameters are explicitly fitted in the reported results.

axioms (1)
  • domain assumption Positive fluctuations correspond to blobs and negative fluctuations correspond to holes
    This mapping is invoked to interpret the dynamics comparison and to label the negative structures as potential artifacts when their behavior deviates from expectation.

pith-pipeline@v0.9.0 · 5721 in / 1456 out tokens · 63008 ms · 2026-05-19T23:39:27.855705+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

17 extracted references · 17 canonical work pages · 1 internal anchor

  1. [1]

    Ben Ayed N, Kirk A, Dudson B, Tallents S, Vann R G L, Wilson H R and the MAST team 2009 Plasma Phys. Control. Fusion 51 035016–40

    work page 2009

  2. [2]

    Plasma 18 060501–49

    D’Ippolito D A, Myra J R and Zweben S J 2011 Phys. Plasma 18 060501–49

    work page 2011

  3. [3]

    Plasma Physics 74 679–717

    Krasheninnikov S I, D’Ippolito D A and Myra J R 2008 J. Plasma Physics 74 679–717

    work page 2008

  4. [4]

    Krasheninnikov S I et al 2003 Proceedings of the 19th IAEA Fusion Energy Conference, Lyon, France, IAEA-CN- 94/TH/4-1

    work page 2003

  5. [5]

    Hasegawa H and Ishiguro S 2017 Nuclear Fusion 57 116008–15

    work page 2017

  6. [6]

    Zweben S J, Terry J L, Stotler D P and Maqueda R J 2017 Rev. Sci. Instrum. 88 041101–22

    work page 2017

  7. [7]

    Fusion 59 056025–40

    Cavalier J, Lemoine N, Brochard F, Weinzettl V , Seidl J, Silburn S, Tamain P , Dejarnac R, Adamek J and Panek R 2019 Nucl. Fusion 59 056025–40

    work page 2019

  8. [8]

    Plasma Phys

    Tamain P et al 2016 Contrib. Plasma Phys. 56 569–74

    work page 2016

  9. [9]

    Han W, Pietersen R A, Villamor-Lora R, Beveridge M, Offeddu N, Golfinopoulos T , Theiler C, Terry J L, Marmar E S and Drori I 2022 Scientific Reports 12 18142–51

    work page 2022

  10. [10]

    Chouche ne S, Brochard F, Desecures M, Lemoine N, and Cavalier J 2024 Scientific Reports 14 27965–80

    work page 2024

  11. [11]

    Chouche ne S, Brochard F, Lemoine N, Cavalier J, Weinzettl V and Desecures M 2024 Phys. Rev. E. 109 045201– 10

    work page 2024

  12. [12]

    Silburn S et al Calcam DOI 10.5281/zenodo.1478554

    work page doi:10.5281/zenodo.1478554

  13. [13]

    Brix M, Hawkes N C, Boboc A, Drozdov V , Sharapov S E and JET-EFDA Contributors 2008 Rev. Sci. Instrum. 79 10F325

    work page 2008

  14. [14]

    Jirakova K, Kovanda O, Adamek J, Komm M and Seidl J 2019 Journal of Instrumentation 14 C11020

    work page 2019

  15. [15]

    Redmon J, Divvala S, Girshick R and Farhadi A 2015 arXiv:1506.02640

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  16. [16]

    Wang C Y, Bochkovskiy A and Liao H Y M 2022 arXiv:2207.02696

    work page arXiv 2022

  17. [17]

    Brochard F et al 2023 Proceedings of the 49th EPS Conference on Plasma Physics, Bordeaux, France

    work page 2023