acceptodds
Withdrawn from ICLR 2027

Physics-Auditable Dryout Recognition in Helical Tubes under Condition Shift

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Reliable dryout recognition from sparse wall thermocouples supports thermal-margin assessment in helical tubes. Existing control-based and raw-channel models can confound condition-dependent temperature levels with circumferential film-redistribution patterns. The main bottleneck is condition transfer in the small-sample regime: pressure and mass flux shift the common temperature level, four-side sampling aliases higher-order circumferential modes, and random splits can obscure group dependence. We introduce a physics-auditable hybrid world-model architecture for retrospective report-label recognition. Its primary lane uses an exact four-point harmonic transform to separate the common level from signed circumferential contrasts before fixed-threshold ExtraTrees classification. An anchored residual RSSM and a differentiable FE decoder provide auxiliary rollout and conditional stress-audit branches. On the MPa pressure-held-out split, the primary lane reaches BA 0.929, compared with 0.786 for raw channels and 0.823 for controls-only ExtraTrees. The five-fold group-held-out BA is 0.957, compared with 0.916 for raw channels. The gain is specific to fixed-threshold BA and does not extend uniformly to AUROC; the focal pressure comparison is not statistically significant. Anchor ablation shows severe rollout drift, although the RSSM remains slightly worse than persistence, and the FE branch remains a conditional structural screen rather than a measured-stress validation. This architecture provides an auditable route from sparse measurements to condition-shifted recognition with explicit evidence boundaries.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Physics-Informed Machine Learning, Hybrid AI Systems, Physics-Anchored World Models, Scientific Machine Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_hGKsOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper investigates dryout recognition in helical tubes using 61 historical experimental runs. Its main method transforms wall-temperature measurements into circumferential averages and contrasts, summarises their evolution over time, and uses an ExtraTrees classifier to predict dryout labels under changing operating conditions. Two alternative representations are explored: to remove shared temperature offsets, and another which uses a recurrent temperature prediction together with a finite-element thermal-stress calculations. The reported classification improvements depend on the evaluation setting, with the main pressure-transfer comparison not statistically significant.

Critical strengths

The underlying ideas are sound, and there seems to be relevant work behind the paper, but the presentation of the work and writing is awful.

Critical weaknesses

Main concern: the quality of the writing is extremely poor and barely readable. The authors use overly complex vocabulary and descriptions for things that should be very simple in essence. It’s missing references almost entirely in the main text and the reader is left to piece things together. Some sentences in the text are just disconnected from any logic or narrative thread. The main figure describing the use case is physically inconsistent, with a pipe that’s actually discontinuous; it is unbelievable that none of the co-authors would notice this! (well at least you don’t need any ML to predict what’s going to happen here: that pipe ain’t ever going to drain or dry out, that’s for sure...). This is a straight “Reject” without hesitation.

Other comments:

Figure-2b: “Summeries” instead of “Summaries”

Which code was used for the FE calculations? The authors mention a “reference implementation” in Appendix L, but these FE calculations remain completely obscure.

Almost no references throughout the main text, with many of the methods like RSSM and ExtraTrees unreferenced.

Some sections are very difficult to decipher, for example Secion-2.3, which describes a “physics-anchored residual world model”. But what is described in that section can hardly be described as “world model”: although the use of this term is not exactly wrong, it is overly grand for what should simply be described as a “predictive recurrent model for the temperature”. Such unnecessary vocabulary amplifications throughout the text make the paper almost impossible to read without significant commitment to actively decipher something that should have been communicated in much simpler terms.

Section-5 Related Work: “Scientific machine-learning methods such as neural operators, DMDc, and physics-informed learning target field-to-field maps and controlled dynamics (Li et al., 2021, Lu et al., 2021, Proctor et al., 2016, Raissi et al., 2019, Karniadakis et al., 2021).” => How does this sentence relate to the rest of the paragraph? That sentence seems to be just stand-alone, not connected to any meaning or logic. The following sentence explains the author’s methods in a nutshell, which are mostly different from these other works. For example, how do PINNs relate in any way to the work presented here?

Questions for authors

see above

Reviewer_8TWzOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies dryout recognition in helical tubes under changing operating conditions. The authors use a harmonic transform of wall temperatures followed by ExtraTrees for classification, with an RSSM and a finite-element model as auxiliary components. Experiments on 61 historical runs show improved recognition under pressure shift.

Critical strengths
  1. The problem is meaningful, and the use of circumferential temperature differences is physically motivated.
  2. The experiments include several useful baselines and ablations. The authors are also careful in reporting limitations, including the lack of statistical significance and the negative results of the auxiliary models.
Critical weaknesses
  1. The evidence for generalization is still limited. The main improvement comes from only 33 held-out runs, with four additional correct predictions and a McNemar p-value of 0.125. The advantage over raw channels also disappears when the pressure transfer is reversed. With only 61 runs from one apparatus, it is difficult to draw strong conclusions about generalization to new conditions.
  2. The contribution of the full world-model architecture is not fully convincing. The main recognition results come from harmonic features and ExtraTrees, while Appendix I shows that simpler contrast statistics can achieve similar performance. The auxiliary RSSM improves rollout stability but does not outperform persistence, and the FE branch remains a conditional stress screen without measured validation. This limits the demonstrated benefit of the complete framework.
Questions for authors
  1. What measurements or physical criteria were used to assign the original dryout labels? In particular, were these labels determined using the same wall-temperature measurements used as model inputs? This seems important given the conflicting label cases in Appendix K.
  2. Could the authors clarify the intended practical benefit of the RSSM and FE branches? The RSSM improves rollout stability but does not outperform persistence, while the FE results are not validated against measured stress.
Reviewer_eHriOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies retrospective recognition of historical dryout report labels from the first 100 seconds of wall temperatures on one helical-tube apparatus. Its primary method applies a fixed four-point circumferential harmonic transform, summarizes the modes over time, and fits ExtraTrees with operating controls. A mean-free covariance extension rejects additive common-mode offsets. Auxiliary branches examine anchored recurrent rollouts and conditional finite-element stress scores. Evaluation uses 61 deduplicated runs with random, pressure, mass-flux, and grouped splits, plus extensive representation, statistical, numerical, and sensor-perturbation audits.

Critical strengths
  • The problem is interesting and has some practical impacts
  • The authors report negative and limiting evidence rather than hiding it.
  • The observation transform has a clear physical interpretation and explicit failure boundaries.
Critical weaknesses
  • The primary recognition method offers limited new learning methodology. The central pipeline combines a fixed four-point Fourier transform, temporal summaries, and an off-the-shelf tree ensemble. The covariance invariances are established projection properties, as acknowledged in the paper. This is a substantive contribution concern for the proposed learning-method framing. The authors should distinguish the useful application finding from an algorithmic advance that the present evidence does not establish.

  • The world-model components do not support the main architectural claim. Table 5 shows that the anchored RSSM remains worse than persistence on every split, and Appendix L reports no independent FE predictive contribution beyond temperature summaries. The structural screen uses late observed fields rather than a validated forecast cascade. These results do not substantiate the benefit of the broader composition; the contribution should be centered on the part that demonstrably helps.

  • The benefit is not specific to the proposed harmonic representation.*The ablations show comparable mean-preserving rotations and substantial classifier dependence; harmonic features can underperform raw features with gradient boosting. The evidence therefore points more narrowly to common-mode separation and contrast fluctuations interacting with ExtraTrees. The authors should avoid interpreting these results as a general superiority of the particular harmonic representation.

  • The incremental transfer effect is statistically uncertain. The focal low-to-high pressure advantage corresponds to only three or four corrected run decisions, and the exact paired tests are not significant. Reverse transfer ties raw features, while the leave-one-cell improvement is one corrected run among 61. Appendix I also leaves the incremental advantage uncertain. This does not prove an absence of benefit, but it is insufficient evidence for a robust general transfer advantage; the conclusion should reflect that uncertainty.

  • External validity and retrospective development remain important limitations. All real-data evidence comes from 61 runs on one apparatus at two pressures, and the focal comparison was developed retrospectively. The breadth of the audits does not create additional independent experimental units. The authors should state which choices preceded inspection of held-out outcomes and limit generalization to the apparatus and conditions actually studied. Broader validation would be future work, not a short rebuttal requirement.

  • The evidence cannot be independently reproduced from the supplied artifacts. Raw data and source code are withheld. Internal filenames and hashes do not permit reconstruction of label provenance, grouping, feature generation, or fitting. This is especially consequential given the small incremental effects and historical report labels. An anonymized per-run prediction and provenance table, if already available, would improve auditability; otherwise, the reproducibility limitation should be explicit.

Questions for authors
  • What specific new learning principle remains after accounting for the known fixed transform and tree readout? What already demonstrated benefit does the RSSM-FE composition add to the primary task, given Table 5 and Appendix L?
  • How were experiment-label groups defined, and which model or representation choices were finalized before inspecting the pressure-held-out outcomes? Please distinguish fixed fitting settings from retrospective development decisions.
  • Is an anonymized per-run prediction table with split, group, and label-provenance identifiers available for independent verification? Please also explain how the main conclusion should be scoped given the reverse-transfer tie and the single-run leave-one-cell increment.

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

Related papers

Discussion (0)

Sign in to comment.