acceptodds
Withdrawn from ICLR 2027

Physics-Aware Heat Partitioning Network

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Multimodal measurements are critical for understanding two-phase boiling heat transfer, yet experimentally resolving phase-specific heat flux contributions remains expensive and challenging. Building on well-established heat-partitioning formulas, we introduce PAPA-Net, a Physics-Aware Heat Partitioning Network for phase-only, mechanism-resolved wall heat flux prediction. PAPA-Net factorizes each heat transfer component into a soft physical support mask and a nonnegative heat-amplitude field, yielding dry area, microlayer, and single phase contributions that sum to the total prediction. Sparse teacher masks are used only during training to assign mechanism identity, while inference requires only optical observations. Physics-coupled objectives enforce consistency between local predictions, mechanism contributions, and integrated heat balance. Across training, validation and external zero-shot conditions, PAPA-Net maintains stable, noncollapsed physical partitions and, on the held-out condition, reduces integrated heat error by 20.4% relative to the next-best optical-only method. Ablations, label-efficiency, robustness, uncertainty, and controlled synthetic experiments further show that the mask-amplitude factorization yields mechanism-aligned heat flux partitions rather than a cosmetic latent decomposition.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
scientific machine learning, physics-aware machine learning, boiling heat transfer, heat flux partitioning, weak supervision, mechanism-resolved learning, multiphase flow, optical-to-thermal inference

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_M8FJOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The proposed method predicts 2D heat flux in boiling conditions, with a separation into dry area, mu layer, and single phase spatial maps, based on masks that are available at training but not inference time. The network produces reasonable separation masks (though their evaluation was very unclear to me) and is able to reduce the overall prediction error vs a single method from 2015. Ablations are carried out. Some key information is missing/unclear. Better fit for publication in application-driven venue.

Critical strengths

The major strength of the paper seems to be the reduced relative energy error in fig 1a combined with the increase in speed reported in table 21. The overall approach seems reasonable; it's unclear if weaknesses below are true to the method or artifacts of poor communication. Note, however, that the contributions of this paper may be a better fit for a venue in the application area rather than a general ML conference.

Critical weaknesses

Very weak comparisons (vs 11 year old baseline and nothing else, key numbers missing e.g. oracle described but dropped from fig 1 and corresponding table, mask-sup dropped from fig 2), ablations suggest that alternatives to the presented method might outperform it on metrics of interest (fig 1 b). Key information missing, particularly the source and meaning (ground truth? helpful guidance?) of teacher masks. Dataset small and details unclear. Extreme use of puffery in language: "x rather than cosmetic y", "a deployable optical-only partition rather than z", etc.

My major gripe is that it was unclear throughout whether the separation into sub-channels was physically accurate or just helpful for the overall prediction, as the overall prediction was the only quantity for which ground truth was reported. The paragraph on line 360 suggests that this division was important, but neglects to mention that seg-reg and mask-sup also do this (and according to fig 1 b in fact do it better). It's also unclear why this wasn't more meaningfully quantified if it was the major contribution of the paper.

Questions for authors

Please address critical weaknesses, particularly around the status of the teacher masks and the primary target of the method.

3 comments

Thank you for the review. We clarify the measured targets, comparisons, and mechanism results using the reported experiments.

Q1. Baselines and figures

Ronneberger et al. (2015) is cited for the U-Net architecture. Figure 1 compares five methods trained on the boiling data, including two supervised segmentation–regression cascades.

Table 1. Held-out S9, three-seed means (Appendix Table 10).

MethodRelE ↓Dice ↑CErr ↓Collapse ↓
Direct U-Net0.0490———
Unguided0.04600.32670.15071.000
Seg-Reg0.05440.84800.01470.000
Mask-Sup0.05380.84800.01380.000
PAPA-Net0.03660.83320.01890.000

PAPA-Net reduces held-out integrated error by 20.4% relative to Unguided, the next-best optical-only method. It runs at 0.97 ms/frame versus 2.09 ms/frame for the cascades (Table 21). The cascades have higher Dice and lower CErr, as Figure 1(b) already shows.

Figure 1 compares optical-only inference. The Fully Guided Oracle requires reference masks at test time and is reported in Figure 2 and Appendix Table 11. Mask-Sup is in Figure 1 and Tables 10–11; we will also include it in Figure 2.

Q2. Primary target and teacher masks

The experimentally reconstructed wall-to-fluid heat-flux field comes from synchronized IR thermometry. Phase-detection imaging identifies dry, microlayer, and liquid-contact regions using optical contrast and interference fringes [1,2]. These masks supervise training; PAPA-Net receives only optical images at inference.

The predicted heat-flux field and total heat transfer are

\qquad \widehat a_k\geq0, \qquad \widehat Q_t=\int_A\widehat q^{\prime\prime}\\,dA.$$ The experimental phase-region references are $$q_{k,\mathrm{ref}}^{\prime\prime}=m_kq_{\mathrm{IR}}^{\prime\prime}, \qquad c_{k,\mathrm{ref}}=\frac{\int_A m_kq_{\mathrm{IR}}^{\prime\prime}\\,dA}{\int_A q_{\mathrm{IR}}^{\prime\prime}\\,dA}.$$ These references quantify heat allocation over optically observed regions (§4.3, Fig. 3). The experiments do not separately measure every microscopic transport process. We will place this definition beside the main results. ### Q3. Mechanism validation Appendix A.7 defines Dice after phase-channel matching and CErr as mean absolute error in integrated contribution fractions: $$\widehat c_k=\frac{\int_A\widehat z_k\widehat a_k\\,dA}{\int_A\widehat q^{\prime\prime}\\,dA}, \qquad \mathrm{CErr}=\frac{1}{3}\sum_{k=1}^{3}|\widehat c_k-c_{k,\mathrm{ref}}|.$$ Active/Collapse measure branch usage. These are reported for all decomposition baselines over three seeds (Table 10). Table 2. S9 controlled ablations, single checkpoint (Tables 14–15). | Configuration | RelE ↓ | Dice ↑ | CErr ↓ | Collapse ↓ | |---|---:|---:|---:|---:| | Full PAPA-Net | 0.0336 | 0.8335 | 0.0180 | 0.000 | | No Factorization | 0.0354 | 0.8373 | 0.0350 | 0.000 | | No Teacher-Mask Guidance | 0.0357 | 0.2836 | 0.2025 | 1.000 | | No Energy Loss | 0.0352 | 0.8339 | 0.0177 | 0.000 | Removing factorization raises CErr by 94.4%. Removing mask-teacher supervision causes complete collapse; supervising 1% of masks restores zero collapse on S9 (Table 16). Removing energy loss increases RelE. The factorization and supervision effects are quantified separately from the five-way baseline comparison. ### Q4. Dataset and relevance to ICLR Table 3. Evaluation coverage. | Dataset | Operating coverage | Protocol | |---|---|---| | dT5 | 7 acquisitions; 0.493–2.172 MW/m² | 10,156 train / 1,792 IID val; 1,948 S9 held-out frames; 7 LOCO tests | | dT10 | 8 heat-flux steps; 48,000 frames | Frozen checkpoints; zero-shot transfer (§4.5) | The learning problem is determining spatial heat allocation from optical observations, measured total heat flux, and sparse phase supervision. Coupled supports, nonnegative amplitudes, and integrated-energy training are tested through the architecture and label-fraction ablations. The ICLR 2027 Call for Papers explicitly includes physical-science applications and physics-informed hybrid AI [3]. We will simplify the language highlighted in the review and present the mechanism definitions alongside Figure 1. References: [1] Kossolapov et al. (2021), *Can LED lights replace lasers for detailed investigations of boiling phenomena?* IJMF. [2] Aguiar et al. (2026), *On the nature of single-phase liquid heat transfer mechanisms in nucleate boiling.* IJMF. [3] ICLR 2027 Call for Papers, https://iclr.cc/Conferences/2027/CallForPapers.

Thank you again for your review. We have uploaded the revised manuscript and updated Figure 2 to include Mask-Sup Cascade.

We also clarified the experimental reference construction, the mechanism metrics, and the five-method comparison. The revised discussion makes the distinction between integrated heat-flux prediction and phase-region alignment explicit, with the corresponding results and ablations.

We hope these revisions address the concerns raised about the evaluation and presentation. Please let us know if anything remains unclear. Otherwise, we would appreciate your reconsideration of the rating.

Thank you for your engagement in the review process. Many of my concerns remain, particularly with regards to clarity and fit to an ML audience, but I have made an important edit to my review: “The contributions of this paper may be a better fit for a venue in the application area rather than a general ML conference,” was included as a weakness because it was a concluding thought resulting from the other weaknesses, not because it was a weakness in itself. The authors are correct that applications papers are included in the ICLR call, but I continue to believe that this specific application paper is a poor fit. To better contextualize my intent with that statement, I am moving it to the strengths section.

The low confidence in the topic area across reviewers, as well as the uniform negativity of reviews, underscores my belief that ICLR is the wrong venue for this work. I hope that it enters the literature through a better-suited peer review mechanism, but I will retain my current score.

Reviewer_bb8QOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PAPA-Net predicts wall heat-flux maps from optical frames and partitions them into dry-area, µ-layer, and single-phase contributions. It combines learned masks with component amplitudes, using sparse teacher masks during training. The paper reports a 20.4% reduction in integrated heat error on a held-out condition and shows that limited mask supervision prevents component collapse.

Critical strengths

The application is relevant, and synchronized optical/IR measurements provide valuable experimental grounding. The ablations also demonstrate an important practical point: accurate total-flux prediction does not guarantee a meaningful decomposition.

Critical weaknesses
  1. Limited novelty and physical justification. The mask–amplitude architecture closely resembles existing masked reconstruction and mixture models. More importantly, the reference components are obtained by masking the measured total flux. This supports phase-region partitioning, but does not establish recovery of distinct heat-transfer mechanisms. The contribution appears primarily to be an optical-only surrogate for an existing experimental analysis pipeline.
  2. The headline improvement is not isolated. The collapsed No-Teacher variant achieves lower held-out integrated error than PAPA-Net. Unguided also removes regularization, so its comparison with PAPA-Net does not isolate teacher guidance or factorization. Without matched objectives and tuning budgets, the reported gain cannot be attributed to the proposed components.
  3. Possible use of the test condition for model selection. Table 8 evaluates loss weights on held-out S9, and Appendix A.6 uses those results to justify the final configuration. This appears inconsistent with the statement that S9 was excluded from model selection. The authors should clarify the selection procedure and provide validation-based tuning results.
  4. The empirical advantage is narrow. PAPA-Net performs worse than the core baselines on MAE, RMSE, and SSIM on both splits, and on IID integrated error. The cascades also achieve better mask and contribution accuracy. Leave-one-condition-out results are provided only for PAPA-Net, leaving the broader comparative claim unresolved. Measurement uncertainty is needed to assess whether the integrated-error differences are physically meaningful.
  5. The synthetic test provides limited independent evidence. Its inputs explicitly contain total-flux information and a blurred component mask, while its targets follow the proposed factorization. Successful recovery under this construction does not establish validity on real optical data.
Questions for authors
  1. Model selection. Was S9 used to select the final loss weights in Table 8? Please report the sweep on the validation split and state the exact selection criterion. How do you reconcile Appendix A.6 with §4.1?
  2. Baseline training. Were Direct U-Net, Unguided, and both cascades trained with the same objective as PAPA-Net (integrated-energy loss, direct head, loss weights, regularizers) and the same tuning budget? If not, please report baselines trained under the identical objective. Please also explain why the collapsed No-Teacher variant and the No-Factorization variant outperform all baselines in held-out RelE.
  3. Teacher masks. How are the dT5 teacher masks generated? If they are obtained by deterministic processing of the phase-detection frames themselves, a non-learned mask extractor could be applied at test time. That would make the “privileged mask” framing and the oracle/cascade comparison very different. If they rely on IR data, please state this explicitly.
  4. Ablation definitions. Please give precise definitions of “No Factorization” and of how “Unguided” differs from “No Teacher-Mask Guidance”.
  5. Measurement uncertainty. What is the measurement uncertainty of the integrated IR heat flux? Is a ~1 percentage-point RelE difference physically meaningful? Why does the Fully Guided Oracle have higher held-out RelE than PAPA-Net?
2 comments

Thank you for the review. We address the five questions with the experimental evidence.

Q1. Model selection

Checkpoint selection minimized IID validation MAE:

e∗(λ)=arg⁡min⁡eMAEIID(e,λ).e^*(\lambda)=\arg\min_e\mathrm{MAE}_{\mathrm{IID}}(e,\lambda).

S9 was excluded from training and checkpoint selection. Appendix A.6 later ranked configurations using S9, with collapse ≤0.10 and active branches ≥2.80 as eligibility criteria. The normalized score weights for Dice/CErr/collapse/activity/MAE/RMSE/SSIM were 0.32/0.28/0.18/0.12/0.04/0.03/0.03. We will correct §4.1/A.6. Five seed-42 checkpoints were evaluated on IID validation alone.

Table 1. Saved checkpoints on IID validation (seed 42).

ConfigurationMAE ↓RelE ↓Dice ↑CErr ↓
Base weak teacher0.036560.018750.84990.02115
Mask weight 0.500.038250.020810.84610.01876
Flux 1.5, direct 0.100.034040.018000.85120.01839
PAPA-Net (final)0.032460.016060.85160.01679
Contribution weight 0.500.035660.018630.84750.01683

The final configuration leads all four IID metrics (MAE −11.2%; RelE −14.3% versus base).

Q2. Baseline training and ablations

All models use the same six sequences, same-seed final-stage splits and energy-loss weight 0.10. Flux/direct weights are 2.0/0.20 for PAPA-Net versus 1.0/0.05 for baselines. Training budgets differ; Tables 14–16 isolate design choices.

Table 2. S9 ablations (single checkpoint, Tables 14–15).

ConfigurationRelE ↓Dice ↑CErr ↓Collapse ↓
PAPA-Net0.03360.83350.01800.000
No Factorization0.03540.83730.03500.000
No Teacher-Mask Guidance0.03570.28360.20251.000
No Energy Loss0.03520.83390.01770.000

The review's S9 ranking is reversed: 0.0336 is lower than 0.0354 and 0.0357. Removing factorization raises CErr by 94.4%; removing mask guidance causes full collapse. With 1% teacher masks, collapse returns to zero (Table 16). Other constraints remain active in these ablations.

Table 3. S9, three-seed means (Table 10).

MethodRelE ↓Dice ↑CErr ↓Collapse ↓
Direct U-Net0.0490———
Unguided0.04600.32670.15071.000
Seg-Reg0.05440.84800.01470.000
Mask-Sup0.05380.84800.01380.000
PAPA-Net0.03660.83320.01890.000

PAPA-Net reduces held-out RelE by 20.4% versus Unguided, with three active branches. Cascades lead on Dice/CErr. On frozen-checkpoint dT10 zero-shot transfer, PAPA-Net has the lowest CErr (0.0783) and collapse (0.166) among optical-only partitioning models.

Q3. Teacher masks and novelty

Optical phase detection supplies dry, microlayer and liquid-contact masks via contrast/interference segmentation [1,2]; synchronized IR supplies heat flux. Optical segmentation is feasible at inference; cascades test learned two-stage segmentation/regression.

\qquad q_{k,\mathrm{ref}}^{\prime\prime}=m_kq_{\mathrm{IR}}^{\prime\prime}.$$ Tables 14–16 quantify how this coupling affects CErr and collapse. The references allocate measured heat across observed phase regions (§4.3, Fig. 3); synthetic components are constructed. ### Q4. Ablation definitions *No Factorization:* Independent nonnegative component maps; phase head retained. *Unguided:* Removes teacher losses and several regularizers. *No Teacher-Mask Guidance:* Removes only the mask-teacher loss; contribution-teacher supervision remains. ### Q5. Measurement uncertainty and oracle Wall-to-fluid energy balance [3,4]: $$q_{\mathrm{w}}^{\prime\prime}=q_{\mathrm{h}}^{\prime\prime}-q_{\mathrm{s}}^{\prime\prime}, \qquad Q_t=\int_A q_{\mathrm{w}}^{\prime\prime}\\,dA.$$ RelE evaluates error in $Q_t$, the heat transferred to the fluid. On S9 it falls from 4.60% to 3.66%: a 20.4% reduction in this physically relevant error. Zhang et al. [3] report ±20 kW/m² applied-flux and ±0.3 °C wall-superheat uncertainties; these do not quantify propagated uncertainty in $Q_t$. All methods use the same IR reference; Tables 10–11 report seed variability and bootstrap intervals. The oracle uses reference masks but still learns heat-flux regression. S9 MAE favors the oracle (0.0329 vs. 0.0422); RelE favors PAPA-Net (0.0336 vs. 0.0538). References: [1] Kossolapov et al. (2021), *Can LED lights replace lasers for detailed investigations of boiling phenomena?* IJMF. [2] Aguiar et al. (2026), *On the nature of single-phase liquid heat transfer mechanisms in nucleate boiling.* IJMF. [3] Zhang et al. (2023), *A unifying criterion of the boiling crisis.* Nat. Commun. [4] Bucci et al. (2016), *A mechanistic IR calibration technique for boiling heat transfer investigations.* IJMF.

Thank you again for your review. We have uploaded the revised manuscript with the additional IID validation comparison, clarified training protocols, and more explicit physical and experimental definitions.

The revision also highlights the controlled ablation results in Tables 14–16, including the held-out RelE comparisons discussed in our response. We have clarified the teacher-mask construction, uncertainty references, and oracle evaluation.

We believe the revised manuscript provides a clearer account of the method and its evidence. Please let us know if any substantive concern remains. We would appreciate reconsideration of your rating in light of these clarifications.

Reviewer_qrqaOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes PAPA-Net, a physics-aware framework for predicting boiling wall heat flux from optical images. The method decomposes total heat flux into dry-area, microlayer, and single-phase contributions through a mask-amplitude factorization. Sparse teacher masks guide mechanism identification during training, while inference requires only optical inputs. Experiments on dT5-Boiling and dT10-Boiling evaluate heat flux prediction, mechanism alignment, robustness, and cross-condition generalization. The results show improved integrated heat flux accuracy under the main held-out condition and more stable mechanism partitioning compared with several optical-only baselines.

Critical strengths
  1. Clear problem formulation. The paper addresses the practical challenge of estimating mechanism-specific heat flux from optical observations. The mask-amplitude factorization provides an interpretable structure, while sparse teacher supervision avoids requiring phase masks at inference.
  2. Comprehensive experimental evaluation. The experiments cover heat flux accuracy, mechanism alignment, ablation studies, robustness, and cross-condition transfer. The analysis also considers branch collapse and performance degradation under low heat flux conditions.
Critical weaknesses
  1. Potential test-set involvement in model selection. The appendix reports hyperparameter experiments on the designated held-out condition and indicates that the final configuration was selected from these results. This raises concerns about whether the held-out evaluation is truly independent and may affect the reliability of the reported generalization performance.
  2. Limited physical validation of mechanism contributions. The mechanism-level reference heat flux is constructed by multiplying the total heat flux by teacher masks, rather than independently measuring each contribution. Therefore, the reported mechanism alignment mainly demonstrates consistency with teacher-derived proxies, not necessarily accurate recovery of the underlying physical contributions.
  3. Limited generalization evidence. The main experiments use data from one facility under restricted operating conditions. The external dT10-Boiling dataset also comes from the same facility, and its relatively high prediction errors suggest that generalization to different experimental setups remains uncertain.
Questions for authors
  1. Was the held-out condition used for hyperparameter tuning or model selection? If so, can the authors provide an evaluation on a completely untouched condition?
  2. How were training and test frames separated within each sequence? Were temporally correlated frames prevented from appearing across different splits?
  3. What independent evidence supports interpreting the teacher-derived heat flux components as physically accurate mechanism contributions?
  4. How robust is PAPA-Net across different facilities or operating conditions, particularly in low-heat-flux regimes where prediction errors and component collapse increase?
3 comments

Thank you for the careful review. We address the four questions with the relevant experimental protocols and results.

Q1. Held-out model selection

Checkpoint selection for each configuration λ\lambda used IID validation MAE:

e∗(λ)=arg⁡min⁡eMAEIID(e,λ).e^*(\lambda)=\arg\min_e\mathrm{MAE}_{\mathrm{IID}}(e,\lambda).

S9 was excluded from training and within-run checkpoint selection. Appendix A.6 later included S9 in cross-configuration ranking with a stability gate; we will clarify this in §4.1 and Appendix A.6. We re-evaluated five saved seed-42 checkpoints using IID validation alone.

Table 1. IID validation of saved configurations (seed 42).

ConfigurationMAE ↓RelE ↓Dice ↑CErr ↓
Base weak teacher0.036560.018750.84990.02115
Mask weight 0.500.038250.020810.84610.01876
Flux 1.5, direct 0.100.034040.018000.85120.01839
PAPA-Net (final)0.032460.016060.85160.01679
Contribution weight 0.500.035660.018630.84750.01683

The reported configuration leads all four metrics among these five, reducing IID MAE by 11.2% and RelE by 14.3% over the base. These results establish validation support for the chosen weights using no S9 metrics in this comparison. Frozen checkpoints were also tested zero-shot on dT10 (§4.5, Appendix B.7).

Q2. Temporal separation

Frames from S4, S6, S7, S8, S11, and S12 were pooled and randomly split 85%/15% into training and validation (seed 42). Sample indices were disjoint, with no temporal exclusion gap; neighboring frames could occur in both partitions. The reported IID test partition uses the validation frames.

S9 is a separate acquisition sequence with no frames in training or IID validation. Its evaluation therefore has no source-frame overlap with the training pool. We will state both protocols explicitly.

Q3. Physical interpretation and mechanism evidence

PAPA-Net imposes nonnegative, phase-conditioned heat-flux partitioning:

\qquad \widehat a_k\geq0.$$ The experimental references combine optically identified phase regions [1,2] with synchronized IR heat flux: $$q_{k,\mathrm{ref}}^{\prime\prime}=m_k q_{\mathrm{IR}}^{\prime\prime}$$ $$c_{k,\mathrm{ref}}=\frac{\int_A m_k q_{\mathrm{IR}}^{\prime\prime}\\,dA}{\int_A q_{\mathrm{IR}}^{\prime\prime}\\,dA}$$ These are experimentally grounded phase-region heat allocations (§4.3, Fig. 3). On S9, three-seed Dice is 0.8332 ± 0.0013, CErr 0.0189 ± 0.0010, with zero collapse (Table 10). Removing factorization nearly doubles CErr (0.0180 → 0.0350); removing mask-teacher guidance causes full collapse, whereas 1% supervision restores zero collapse (Tables 14–16). The synthetic study additionally tests recovery against known constructed components. We will distinguish these measurements from independently measured microscopic mechanism fluxes. ### Q4. Cross-condition and zero-shot robustness Seven leave-one-condition-out (LOCO) experiments exclude the evaluated condition from training (Table 18). Five have zero collapse; six are below 0.10. At 0.493 MW/m², RelE is 1.120 and collapse 0.926, with dry events present in only 2.85% of frames. Integrated heat error is evaluated as $$\mathrm{RelE}=\frac{1}{N}\sum_{t=1}^{N}\frac{|\widehat Q_t-Q_t|}{|Q_t|+\varepsilon}, \qquad Q_t=\int_A q_{\mathrm{IR}}^{\prime\prime}\\,dA.$$ The denominator makes relative error particularly sensitive when the reference integrated heat is small. Table 2. S9 held-out (three-seed means) and dT10 zero-shot (eight steps, 48,000 frames). | Method | S9 RelE ↓ | dT10 RelE ↓ | dT10 CErr ↓ | dT10 Collapse ↓ | |---|---:|---:|---:|---:| | Direct U-Net | 0.0490 | 0.5358 | — | — | | Unguided | 0.0460 | 0.5104 | 0.1675 | 0.800 | | Seg-Reg | 0.0544 | 0.6434 | 0.0903 | 0.176 | | Mask-Sup | 0.0538 | 0.6426 | 0.0932 | 0.172 | | PAPA-Net | 0.0366 | 0.5320 | 0.0783 | 0.166 | S9 RelE improves by 20.4% over Unguided. dT10 is an intentionally stringent zero-shot stress test: dT5-trained checkpoints are frozen and evaluated under changed subcooling and acquisition conditions, without target-domain training or calibration. PAPA-Net retains the lowest CErr and collapse among optical-only partitioning methods; the reported RelE characterizes transfer under that shift. Both datasets use the same facility. We will specify this evaluation scope in the revision. References: [1] Kossolapov et al. (2021), *Can LED lights replace lasers for detailed investigations of boiling phenomena?* IJMF. [2] Aguiar et al. (2026), *On the nature of single-phase liquid heat transfer mechanisms in nucleate boiling.* IJMF.

Thank you again for your review. We have uploaded a revised manuscript incorporating the clarifications discussed in our rebuttal. In particular, §4.1 and Appendix A now document the selection and splitting protocols, including the additional IID validation results. We also clarified the physical interpretation of the reference components and the scope of the LOCO and zero-shot evaluations.

We believe these revisions address the principal questions raised. Please let us know if any concern remains. Otherwise, we would appreciate your reconsideration of the rating.

I thank the authors for the detailed and unusually candid response. I keep my score unchanged.

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

Related papers

Discussion (0)

Sign in to comment.