acceptodds
Under review as a conference paper at ICLR 2027

ScaleComposer: Scale-Conditioned Compositional Learning for Source-Count Extrapolation in Physical Fields

Abstract

Air movement enhances convective heat exchange with occupants and exposed surfaces, motivating accurate prediction of spatial air speed distributions in buildings. As the number of interacting sources grows, characterizing airflow across all possible configurations becomes costly. Predicting collective airflow responses from limited measurements is therefore critical. Neural operators can capture complex interactions and multiscale features by learning mappings between function spaces. However, representative architectures do not explicitly represent interactions among momentum sources or their dependence on spatial scale. We propose a compositional learning framework that builds reusable representations of individual sources and source pairs from single-source priors and inter-source relations. Location- and scale-conditioned aggregation combines these representations, allowing the same learned composition rules to operate across source counts. Residual reconstruction then refines a pretrained reference field by constraining scale-specific corrections to corresponding spatial frequency bands. This connects the scale at which source information is aggregated to the spatial structure of the predicted corrections. We evaluate extrapolation from single- and two-source configurations to configurations with up to six sources, using measured multi-ceiling-fan airflow and six controlled numerical regimes spanning source overlap and advective dynamics. On measured data, the framework reduces overall extrapolation error by up to 17.1% relative to the baselines. Replacing adaptive aggregation with fixed pooling improves low-source validation performance but worsens extrapolation, highlighting the importance of learning how source information is combined. In controlled experiments, the framework demonstrates competitive overall error and better reconstructs spatial patterns than baselines.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
scientific machine learning, compositional generalization, neural operators, multiscale modeling, relational representation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_Dbd1OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper addresses the challenge of predicting time-averaged scalar fields, such as ceiling-fan air speed, for settings where there are more active sources (e.g. fans) than seen in the training set. The authors propose a new model, ScaleComposer, which combines individual fan and fan-pair tokens through separate attention modules (conditioned on location and band). A decoder decodes residual corrections to a frozen reference field with a shared DeepONet and those outputs are finally projected onto fixed Fourier frequency bands as the final output. The approach is evaluated on multi-fan airflow (three test settings with 3–6 fans) and a 2D Navier–Stokes benchmark. The authors report improved performance over baselines on the multi-fan airflow experiment.

Critical strengths

I want to highlight the following strengths of the submission:

  • I found the paper generally well written and easy to follow; it's structure well, the methods section is intuitive and not over-complicated / over-mathy; the related work section touches on relevant literature.
  • I thought the figures are generally pretty nice and do a good job at conveying the main contributions and experimental results visually.
  • As far as I can tell, the proposed approach is quite interesting and methodologically novel (and it is well suited to the problem setting)
  • While the problem setting is a bit more niche, it is motivated and presented well.
Critical weaknesses

The submission has the following shortcomings, ordered roughly from, in my opinions, most to least severe:

  • Unless I got something wrong, the main conclusion (ScaleComposer beating existing methods on a multi-fan airflow setting) relies on findings on one single and quite small test set. It's nice to see confidence intervals at least, but I am still not sure how generalizable the finding is.
  • I found some ablations missing that I was wondering about: (1) It might be cool to do a fan-only ablation to see the difference to the fan + pair variant; curious why the authors did not try that. (2) Similarly, I am wondering why the authors did not compare their sigmoid with the 1 + G denominator to a standard softmax?
Questions for authors

Overall, this is a decent submission. A bit more niche application-wise, but an interesting new methodology. I am less convinced by the lack of experiments and some lack of ablations, but overall, this is still an interesting submission.

For questions; please elaborate on the ablations (see shortcomings).

1 comment
AuthorsOpenReview ↗

Response to Reviewer Dbd1

We appreciate the positive and constructive comments from you! We have modified our paper according to your comments.

W1&Q1: the main conclusion (ScaleComposer beating existing methods on a multi-fan airflow setting) relies on findings on one single and quite small test set. It's nice to see confidence intervals at least, but I am still not sure how generalizable the finding is.

We appreciate this concern and agree that evaluation within a single physical environment limits the generalizability that can be established experimentally. Please refer to the general response 2.

W2&Q2: some ablations missing that I was wondering about: (1) It might be cool to do a fan-only ablation to see the difference to the fan + pair variant; curious why the authors did not try that. (2) Similarly, I am wondering why the authors did not compare their sigmoid with the 1 + G denominator to a standard softmax?

Thank you for these suggestions. We have added both ablations. Table 8 details their constructions, while Tables 9 and 10 report their results on the measurement and numerical datasets, respectively. We also discuss these findings in general response 3.1 and have updated Section 5. Results and Analysis of the manuscript accordingly.

(1) Fan-only's difference to fan + pair variant

In the original submission, we evaluated the combined source–pair representation without separately isolating the contribution of pair tokens. We agree that this comparison is important for understanding the role of explicit pairwise representations, and this omission made it difficult to assess its contribution. As stated in general response 3.1.1, we have therefore added a SourceOnly variant that retains fan attention but removes the pair representation. On measurement, its MAE rises from 0.180 to 0.205 m/s (+13.9%) and the PCC falls from 0.798 to 0.745. The across-seed standard deviation of MAE also increases from 0.009 to 0.025 m/s. On the numerical benchmark, PCC decreases from 0.871 to 0.849. These results indicate that explicit pairwise representations improve prediction accuracy and stability. While fan tokens represent individual sources, pair tokens additionally encode relative geometry and operating conditions between sources, providing information relevant to their interactions.

(2) "Compare their sigmoid with the 1 + G denominator to a standard softmax"

We have added a variant using standard softmax over valid items. As stated in general response 3.1.2, its measurement MAE increases from 0.180 to 0.188 m/s, while its numerical PCC decreases from 0.871 to 0.851. The motivation for our normalization is particularly relevant to extrapolation across source counts. For two-fan training samples, there is only one valid pair token. Standard softmax therefore assigns it a weight of one regardless of its attention score, providing no gradient through the attention weighting for that single pair. At test time, however, configurations with up to six fans introduce as many as fifteen pair tokens, whose softmax weights must share a fixed total of one.

In contrast, our sigmoid-based normalization with the 1 + G denominator retains score-dependent weighting even for a single pair and does not constrain the aggregate weight of valid pair tokens to equal one. This allows the model to adjust the contribution of pairwise information as the number and relevance of source pairs change. The observed improvements support this normalization choice for extrapolation to larger source configurations.

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

This paper investigates source-count extrapolation in physical-field prediction. The goal is to train a model using configurations with at most two active sources and predict the airflow fields generated by configurations containing three to six sources. The authors propose ScaleComposer, which first constructs an initial reference field using pretrained single-source predictions and a composition model. It then learns corrections to this field by representing individual sources and their pairwise interactions. Specifically, separate scale-conditioned attention modules aggregate source and source-pair features, while a shared DeepONet decoder generates residual corrections constrained by predefined Fourier frequency bands. These corrections are combined with the reference field to produce the final predicted air-speed distribution. The method is evaluated on two datasets: real-world measurements from a room equipped with six ceiling fans and a controlled numerical benchmark generated using two-dimensional Navier–Stokes simulations under six different physical regimes. On the measurement dataset, ScaleComposer reduces MAE by approximately 10% compared with the strongest evaluated baseline. However, on the numerical benchmark, the full model underperforms several baselines, while removing the output reference addition substantially improves its accuracy.

Critical strengths

S1. Meaningful and practically relevant problem.

The paper addresses a meaningful generalization problem in scientific machine learning. Learning from low-source configurations to predict physical fields under larger source sets has practical relevance to airflow modeling and other multi-source physical systems.

S2. Carefully designed numerical experiments.

The numerical benchmark systematically varies source overlap and advective influence, including matched advection-off controls. This design helps distinguish nonlinear transport effects from the non-additivity introduced by scalar-speed observations.

S3. Detailed implementation and experimental analysis.

The paper provides comprehensive implementation details, including model architectures, training procedures, numerical verification, and computational costs. The combination of real-world measurements and controlled numerical experiments, together with architectural ablations, is valuable for evaluating the proposed framework.

Critical weaknesses

W1. Insufficient statistical evidence for robust source-count extrapolation.

A major concern is the limited statistical support for the real-world evaluation. According to Appendix Table 2, the test set contains only three higher-source configurations, corresponding to three, four, and six active fans. Each source count is represented by only one configuration, while source locations, operating conditions, and pairwise interactions also differ across these configurations.

Consequently, the observed performance differences cannot be attributed solely to increasing source count. The evaluation also does not establish whether the proposed method can consistently generalize across different configurations containing the same number of active sources.

Furthermore, the reported standard deviations are computed across five random training seeds rather than independent test configurations. Although these results characterize training variability, they do not capture uncertainty arising from variations in physical configurations. The approximately 10% MAE improvement over the strongest baseline (0.183 vs. 0.203 m/s) is promising, but its robustness across independent configurations remains unclear.

This limitation is particularly relevant because different configurations were measured in separate sessions, and the measurements themselves are subject to instrumental uncertainty. Although the numerical experiments provide broader configuration coverage, they cannot fully compensate for the limited statistical evidence from the real-world evaluation.

W2. Limited empirical support for the proposed scale-conditioned compositional mechanism.

The paper's main methodological contribution lies in combining separate source/pair attention with spectrally constrained residual reconstruction. However, the controlled numerical experiments provide limited evidence that these components consistently improve source-count extrapolation.

According to Appendix Table 8, separate attention achieves an MAE of 0.070 compared with 0.071 for joint attention, indicating only a marginal improvement. Similarly, removing spectral band restrictions slightly improves MAE from 0.070 to 0.069. Although the authors report improvements in other metrics and frequency components, the overall MAE results do not demonstrate a consistent advantage for these two central design choices.

Furthermore, the scale-resolved analysis indicates that 97.7% of the six-source MSE reduction relative to the reference field comes from the DC and lowest-frequency components. While this observation is informative, it does not establish that the proposed scale-conditioned relational composition is necessary to obtain these improvements. Simpler global or low-frequency correction models might achieve similar benefits.

The measurement-domain ablations suggest that separate attention and spectral constraints improve performance, but these results are based on only three higher-source configurations. Consequently, the strongest evidence supporting the proposed architectural components comes from the statistically limited part of the evaluation.

Overall, the current experiments do not sufficiently isolate the contribution of scale-conditioned composition from simpler alternatives, such as low-frequency residual correction or general spectral regularization.

W3. Underperformance of the full model on numerical benchmarks and unclear model-variant selection.

Another major concern is the inconsistency between the claimed effectiveness of the framework and the numerical benchmark results.

According to Appendix Table 6, the full ScaleComposer model achieves an MAE of 0.070, which is substantially worse than several competing methods, including DeltaPhi-DeepONet (0.047), Ref+GNOT (0.053), Ref+FNO (0.054), and Ref+GINO (0.054).

Removing output reference addition improves ScaleComposer's numerical MAE to 0.047, matching DeltaPhi-DeepONet. However, this variant does not consistently outperform the baselines across all metrics. For example, its PCC of 0.798 remains below Ref+GNN (0.850) and Ref+GNOT (0.836).

These results raise an important concern regarding the general effectiveness of the proposed framework. The full reference-assisted model performs best on real-world measurements, whereas removing output reference addition substantially improves its performance in the numerical experiments. Importantly, the reference-free output variant only removes the final addition of the reference field; reference-related information remains available to the model. Thus, this ablation evaluates the contribution of output reference addition rather than completely removing prior information.

Although the paper acknowledges this domain-dependent behavior, it does not establish a clear criterion for selecting the appropriate model variant before evaluating higher-source test configurations. If the choice between variants is informed by their performance on higher-source test sets, the reported results may reflect test-dependent model selection rather than a predefined, reproducible prediction framework. This does not establish that such selection occurred, but the selection protocol should be explicitly clarified.

Without a validation-based selection criterion or a systematic analysis of reference-field quality, it remains unclear under which conditions the full ScaleComposer model can be expected to outperform existing baselines.

Questions for authors

Q1. Reliability of the Real-World Evaluation.

The real-world evaluation contains only three higher-source configurations, with one configuration for each evaluated source count. Could the authors provide configuration-level and height-wise results, together with an uncertainty analysis, to better substantiate the approximately 10% MAE improvement? How can the effects of source-count extrapolation be distinguished from simultaneous changes in source layouts and operating conditions?

Q2. Effectiveness of the Core Architectural Components.

In the controlled numerical experiments, separate attention provides only marginal MAE improvement, while removing spectral band restrictions slightly improves MAE. Furthermore, most of the MSE reduction comes from the DC and lowest-frequency components. Could the authors provide more direct evidence that scale-conditioned relational composition is necessary, for example through comparisons with simpler global or low-frequency residual correction models and sensitivity analyses of the frequency-band design?

Q3. Numerical Performance and Model-Variant Selection.

The full ScaleComposer model underperforms several baselines on the numerical benchmark, whereas removing output reference addition substantially improves its MAE. How was the choice between these two variants determined? Could the authors provide a validation-based selection criterion and explain under which conditions reference addition is beneficial, without relying on higher-source test results?

3 comments
AuthorsOpenReview ↗

Response to Reviewer MruM (Part 1/3)

We sincerely thank Reviewer MruM for the constructive comments!

1.1 Response to W1

We thank the reviewer for this detailed comment. Please see General Response 2 for the expanded measurement evaluation. Below we explain how it addresses each part of this concern.

1.1.1 Number of configurations with the same source count.

We agree that one configuration per source count was insufficient. We have added two six-fan configurations, so the test set now contains three six-fan cases (General Response 2.1). They share the same source layout but with different operating conditions (all fan operating at 20% rotational speed, all fan operating at 100% rotational speed, mid-row two fans operating at 20% rotational speed with other four operating at 100%). Even with the same source count and fan layout, wide MAE range across the three six-fan cases (0.09 m/s to 0.262 m/s) confirms that performance differences across configurations cannot be attributed solely to source count. The three- and four-fan counts are still represented by one configuration each.

1.1.2 Variability across configurations. We agree that seed-based standard deviations do not capture configuration variability. Table 2 now reports results separately for each test configuration. ScaleComposer reduces MAE relative to Ref+GNN in all five configurations, by 3% to 13%. The height-wise standard deviations in Table 2(a) are computed over configurations and seeds, and therefore include configuration variability. With the expanded test set, the overall reduction is 8% (0.180 vs. 0.196 m/s). Five configurations are still a small sample. We therefore report the per-configuration results directly, so that consistency can be judged case by case.

1.1.3 Measurement sessions and instrumental uncertainty. All configurations followed the same measurement protocol. At each probe location, a 30 s stabilisation period was followed by 100 s of sampling at 1 Hz. The reported value is the mean of these samples. At 80 s, the running mean at every point lies within ±6.74% of its final 100s mean, and the spatially averaged running mean lies within 0.01%. The reported means are therefore insensitive to the averaging window. In addition, all models are evaluated against the same measured fields. Session effects and instrumental errors are thus shared across models. Because each configuration was measured in a separate session, the consistent improvement across all five configurations also suggests that the advantage is not tied to a single session.

1.1.4 Role of numerical experiments. We agree that numerical experiments cannot fully compensate for limited measurement data. Here we use them as complementary evidence with broader configuration coverage. On both datasets, ScaleComposer remains below the reference field at every source count and has the lowest MAE among the compared models at four or more sources (general response 2.4). We have stated the small number of measured configurations as a limitation in Section 5.2.

1.2 Response to Q1

This question overlaps with W1, and our response to W1 gives the details. We summarize the answer to each part here.

1.2.1 Configuration-level and height-wise results. Table 2 now reports MAE by test configuration and by measurement height for ScaleComposer and Ref+GNN (General Response 2.2). ScaleComposer has lower MAE in all five configurations (3%–13%) and at all four heights (2%–11%). With the expanded test set, the overall MAE reduction is approximately 8% now.

1.2.2 Variability and measurement uncertainty. For configuration variability, Table 2 reports MAE separately by configuration and by height. For each configuration, the standard deviation is computed across heights and seeds; for each height, it is computed across configurations and seeds. These statistics describe variation across the corresponding combinations. Training variability was averaged over five seeds. For measurement uncertainty, each point is a 100 s mean obtained after a stabilisation period. Details are given before in our response to W1.

1.2.3 Separating operating conditions from layout and source count The three six-fan test cases in the measurements fix both source count and layout, and vary only fan speeds. Their MAE varies widely, which suggests that operating conditions alone can strongly affect prediction difficulty. ScaleComposer still improves on Ref+GNN in each of them. For the effect of source count, we therefore rely on the numerical benchmark, which contains multiple layouts at each source count. General Response 2.4 shows that the MAE of ScaleComposer remains below the reference field at every source count and is the lowest at four or more active sources compared with other baselines.

AuthorsOpenReview ↗

Response to Reviewer MruM (Part 2/3)

2.1 Response to W2

We thank the reviewer for this careful analysis. Please see General Response 1 for the reference-field correction, General Response 3.1 for the expanded ablations, and General Response 4.2 for the band-boundary sensitivity analysis. Below we explain how they address each part of this concern.

2.1.1 Numerical results for separate attention and band restriction. The MAE values cited here (0.070, 0.071, and 0.069) were obtained with the erroneous additive reference. With the corrected one, the full model achieves the lowest MAE and RMSE and the highest PCC among all ablation variants. Single joint attention increases MAE to 0.035 and lowers PCC from 0.871 to 0.852. Removing band restriction no longer improves MAE. It increases MAE to 0.035 and lowers PCC to 0.846. We agree that the MAE differences remain small on the numerical benchmark. The clearer difference is in PCC, which reflects whether the spatial structure is reproduced.

On measurements, the differences are larger on both metrics. Single joint attention and w/o band restriction increase MAE and lower PCC. One possible reason for the benefit of separate attention lies in how the item counts change. Each training sample contains two sources and one pair, whereas six sources give fifteen pairs. Single joint attention shares projections and a common normalization across both item types, so the more numerous pair items may dominate the aggregated output. Separate attention aggregates sources and pairs separately.

2.1.2 Comparison with simpler global or low-frequency correction. We agree that this comparison was missing. General Response 3.1.3 adds three such variants: a global constant correction (DC-only), a low-frequency correction (DC and B1), and a low-frequency correction that uses mean pooling instead of relational attention. None reaches the MAE of the full model on either dataset. More importantly, they barely improve the spatial pattern of the reference field. On measurements, their PCC falls below that of the reference (0.741–0.774 vs. 0.785). On the numerical benchmark, it stays close to the reference (0.843–0.849 vs. 0.844). The full model raises PCC to 0.798 and 0.871, respectively.

2.1.3 Interpretation of the scale-resolved analysis. After the correction, the DC and lowest-frequency components account for 84.7% of the six-source MSE reduction, rather than 97.7%. We agree that they still account for most of it. These components correct the overall magnitude, which likely dominates squared error. The higher bands contribute less to MSE, but they appear necessary to correct the spatial pattern. A low-frequency correction alone can therefore reduce MSE substantially without reproducing the spatial structure.

2.1.4 Spectral design. General Response 4.2 scales the internal band boundaries by 0.8 and 1.2. Measurement MAE stays within 0.179–0.181 m/s. On the numerical benchmark, the original partition performs best, and accuracy is more sensitive to shifting the boundaries toward lower frequencies. The benefit is therefore not tied to a finely tuned partition, although the boundaries should not be shifted too far toward lower frequencies.

2.1.5 Statistical basis of the measurement ablations. The measurement test set now contains five configurations (General Response 2.1; see also our response to W1). With the corrected numerical results, both datasets show the same ranking: the full model performs best among the ablation variants on all three metrics, with one tie in measurement PCC. The numerical benchmark therefore no longer contradicts the measurement ablations, although its differences are smaller and appear mainly in PCC. Figure 3(c, d) also breaks down the ablations by source count. Among the variants shown, the full model has the lowest MAE at every source count except two. On measurements at three sources, w/o scale conditioning is slightly lower. On the numerical benchmark at five sources, SourceOnly is slightly lower.

2.2 Response to Q2

Marginal numerical MAE differences. The cited values came from the erroneous additive reference. With the corrected reference, the full model performs best among all ablation variants, and removing band restriction no longer improves MAE. The numerical advantage appears mainly in PCC (General Responses 1 and 2.3).

Simpler global or low-frequency correction. We added DC-only, Low-frequency, and Low-frequency (pooling) variants (General Response 3.1.3). They reduce error relative to the reference field but barely improve its spatial pattern, while the full model does. This explains why most (84.7%) of the MSE reduction comes from low frequencies, while the higher bands are still needed for spatial accuracy.

Sensitivity of the frequency-band design. We scaled the internal band boundaries by 0.8 and 1.2 (General Response 4.2). Measurement MAE stays within 0.179–0.181 m/s. On the numerical benchmark, the original partition performs best.

AuthorsOpenReview ↗

Response to Reviewer MruM (Part 3/3)

3. Response to W3 and Q3

We sincerely thank the reviewer for the detailed evaluation of the numerical benchmark, which led us to find the error described in General Response 1. We answer the concern by three main parts below.

3.1 Numerical performance of the full model. The MAE of of 0.070 was obtained with an additive reference used by mistake in the numerical pipeline (detailed in General Response 1). The “w/ Additive Ref” variant in Table 10 reproduces the same value. All reference-based baselines were also rerun with the corrected reference. With the intended non-additive reference, the full model reaches an MAE of 0.033, which is now better than the competing methods in Table 7. Its PCC (0.871) is also higher than that of Ref+GNN (0.855) and Ref+GNOT (0.845). It ties POD-RBF for the lowest MAE and Ref+GNN for the lowest RMSE, and has the highest PCC. By source count, it is the most accurate model at four or more sources.

3.2 The reason why removing the output addition helped in previous manuscript. The previous erroneous reference field used on numerical benmarks sumed up single-source predictions and background field. It ignores the interaction between sources, and its error tends to grow as sources are added. Furthermore, the residual part was trained on two sources and is therefore difficult to correct an error that grows beyond that range. Removing such field from the output accidentally avoid it.

Besides, the original “ScaleCompsoer w/o Ref” ablation only removed reference field from the output addition. We have removed it from both the input and the output in the present version. This update slightly reduces its mean MAE from 0.047 to 0.046 on the numerical benchmark but increases from 0.301 to 0.307 m/s on the measurements.

3.3 When the reference addition is beneficial.

With the corrected reference, the full model is the best variant in both domains, so choice on reference field usage is actually no longer required. The non-additive reference is the intended design already used for the measurements. Using the additive reference is worse than removing the reference altogether.

The table below shows the mean MAE of additive reference field and non-additive referenc field on two-source validation set. The results indicate the importance of reference field quality in the proposed method that can be further validated in the future work.

DatasetAdditive Reference MAENon-additive Reference MAE
Measurement (m/s)0.4370.268
Numerical (dimensionless)0.0310.022
Reviewer_dZbwOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes ScaleComposer, a framework for predicting physical fields with more active sources than those observed during training. The method combines source-level and pairwise representations using separate scale-conditioned attention modules, followed by spectrally constrained residual corrections to a pretrained reference field. The authors evaluate the method on measured ceiling-fan airflow and six controlled Navier–Stokes simulation regimes, training on configurations with at most two sources and testing on configurations with three to six sources. Results show approximately 10% lower MAE than the strongest baseline on measured data, with ablations supporting the proposed attention and frequency-band components. Overall, the paper addresses an interesting compositional generalization problem and presents promising results.

Critical strengths
  1. The experiments show that ScaleComposer outperforms several graph-based, attention-based, and Fourier-based baselines on measured airflow data. The improvement is relatively consistent across different source counts, and the ablation studies suggest that separate source–pair attention and spectral constraints contribute to performance.

  2. I appreciate the additional experiments on controlled Navier–Stokes systems, especially the investigation of source overlap and nonlinear advection. The scale-wise error analysis also provides useful insights into which spatial frequencies contribute most to the improvement.

Critical weaknesses
  1. Although the paper compares against several established neural operators, it is unclear whether the proposed architectural complexity is necessary. In particular, comparisons with simpler source-wise or pairwise aggregation, combined with the same pretrained reference and residual decoder, would be useful. I would also like to see a baseline that learns only low-frequency or DC corrections, since the analysis shows that 97.7% of the six-source MSE reduction relative to the reference comes from the mean and coarsest frequency components. Such comparisons would better isolate the value of the scale-conditioned attention and spectral reconstruction.

  2. The paper motivates scale-conditioned compositional learning, but the improvements are not consistent across the two experimental domains. In particular, the numerical experiments show limited benefits from separate attention and mixed results for band restrictions across different metrics. This makes it unclear whether scale conditioning provides a general advantage or primarily improves the measured airflow task. Furthermore, despite having fewer than one million parameters, ScaleComposer has relatively high computational cost, and its reported inference latency excludes the construction of pretrained reference fields. A clearer accuracy–computation comparison would strengthen the practical contribution.

Questions for authors

I would also like to see comparisons with classical reduced-order modeling (ROM) approaches, which are widely used for efficient CFD prediction and would serve as important baselines for evaluating the proposed method.

2 comments
AuthorsOpenReview ↗

Response to Reviewer dZbw (Part 1/2)

Thank you for your insightful and constructive comments! We have added additional ablations and baseline, as well as modified our paper according to your comments.

W1: Comparisons with simpler source-wise or pairwise aggregation, combined with the same pretrained reference and residual decoder, would be useful. I would also like to see a baseline that learns only low-frequency or DC corrections.

We now include the comparisons suggested. Please see full results in General Response 3.1.1 Representation and poolingand 3.1.3 Residual reconstruction, respectively. Adding pair features to uniform pooling does not improve measurement MAE, whereas incorporating them into the full attention design improves upon SourceOnly. These results support the combination of explicit pair representations and adaptive aggregation beyond the shared reference and decoder. Besides, even the strongest restricted variant does not match the full model. The degradation observed when DC is removed further supports combining mean and spatially varying corrections.

W2 (1) The improvements are not consistent across the two experimental domains. In particular, the numerical experiments show limited benefits from separate attention and mixed results for band restrictions across different metrics. This makes it unclear whether scale conditioning provides a general advantage or primarily improves the measured airflow task.

Please see General Response 1 for the reference-field correction, General Response 2.3 for the updated comparisons on numerical benmarks and General Response 2.4 for the source-count comparison on both datasets. Below we explain how they address the concern.

We agree that the original results did not show a consistent advantage across two domains. Part of this inconsistency was caused by the reference error in the numerical pipeline. With the corrected reference, the full ScaleComposer gives the lowest error on both datasets. Its advantage over baselines also follows a similar pattern in both domains. It remains below the reference field at every source count and is the most accurate of the compared models at four or more sources. Its advantage is therefore not specific to the measured airflow task.

Individual components contribute differently in the two domains. On the numerical benchmark, the component variants differ from the full model mainly in spatial pattern agreement, not in pointwise error. We see this as a difference in what the metrics capture, rather than a contradiction. MAE and RMSE reflect local magnitude errors, while PCC reflects whether the spatial structure is reproduced. Scale conditioning and band restriction appear to act mainly on the spatial structure. One plausible reason for their smaller effect on pointwise error is that the learned reference already captures much of the numerical field. This leaves a smaller residual for these components to correct.

W2 (2) ScaleComposer has relatively high computational cost, and its reported inference latency excludes the construction of pretrained reference fields. A clearer accuracy–computation comparison would strengthen the practical contribution.

We agree that the end-to-end latency should be reported for all Ref+ methods and for ScaleComposer. Table 6 now includes reference field evaluation (4.87 ms) in the inference latency. The construction of pretrained reference field is a one-time offline cost shared by all Ref+ methods. Furthermore, Figure 6 provides a clearer accuracy-computation comparison for the proposed model and baselines.

We agree that ScaleComposer is not the cheapest model. Its FLOP count (14.548 G) is much higher than that of Ref+FNO (0.171 G) and Ref+GNN (4.443 G). However, this does not translate into a proportional runtime cost. Compared to Ref+FNO, the fastest reference-based baseline, its end-to-end latency is only 1.7 times higher (13.507 ms against 8.022 ms), and its training throughput is about half (0.492 against 0.995 epochs/s). Compared with Ref+GNN, the baseline with the lowest MAE, latency increases by 2.6 ms (13.507 ms against 10.948 ms) and throughput falls from 0.637 to 0.492 epochs/s. All reference-based methods except Ref+GINO (157.519 ms) fall within 8.0–13.7 ms per plane. For this cost, ScaleComposer attains the best result on all three metrics. Its MAE is 0.180±0.009 m/s, against 0.196±0.009 for Ref+GNN and 0.241±0.011 for Ref+FNO, a reduction of 8% and 25% respectively.

The two baselines without a reference field are faster but less accurate. POD-RBF has the lowest latency (0.138 ms) owing to its offline–online decomposition, but its MAE on measurements is more than twice that of ScaleComposer. DeltaPhi-DeepONet (5.132 ms) shows a similar trade-off. Overall, ScaleComposer’s accuracy gain costs a few milliseconds per plane. This remains small compared with a single CFD simulation of the same configuration, ranging from 3 days to over a week.

AuthorsOpenReview ↗

Response to Reviewer dZbw (Part 2/2)

Q: I would also like to see comparisons with classical reduced-order modeling (ROM) approaches.

We have added a POD-RBF baseline as a classical ROM approach. Please refer to the general response "3.3 Classical reduced-order modeling (ROM) baseline" for details.

In brief, POD-RBF is competitive on the numerical benchmark and matches ScaleComposer in aggregate MAE. The source-count breakdown shows a clearer difference. POD-RBF is more accurate at three sources, while ScaleComposer is more accurate at four, five, and six sources. On measurements, ScaleComposer achieves lower MAE than POD-RBF at every evaluated source count. POD-RBF has much lower inference latency because of its offline–online decomposition. However, its POD basis is derived from training snapshots. It may not represent the new spatial structures that arise when more jets interact. In our tests, POD-RBF is therefore efficient and accurate near the training range, while ScaleComposer generalizes better to larger source counts and to measured data.

Comments on the paper

3 comments
AuthorsOpenReview ↗

General Response to All Reviewers (Part 1/3)

We express our gratitude to all the reviewers for their perceptive comments aimed at enhancing the quality of our work.

1. Correction to the reference field in the numerical experiments

We sincerely thank the reviewers for the careful examination of the manuscript, which prompted us to revisit the reference-field pipeline. In doing so, we identified an error in the numerical experiments: the numerical pipeline used an additive reference instead of the intended learned non-additive reference. We apologize for this oversight. As a result, the numerical results in the original submission were obtained with the additive reference. We have rerun all affected numerical experiments with the learned nonadditive reference, keeping all other settings (data splits, hyperparameters, and training protocol) unchanged, as well as updated related contents in abstract, Section 5, Section 6, and Appendix accordingly. The measurement experiments used a separate, correctly configured reference pipeline and are not affected.

The additive reference was constructed during development from the available vector-valued single-source predictions, to assess the importance of reference-field construction quality. It was not included in the original ablation study because its measurement counterpart, based on scalar single-source predictions, had not been implemented. We now report it as an additional ablation (please see construction details in Table 8) for both numerical experiments (in Table 10) and real-world measurement experiments (in Table 9) . The corrected comparison gives an MAE of 0.033 with the learned non-additive reference, versus 0.070 with the additive reference. This suggests that reference quality is an important contributor to performance.

We also agree with the suggested revision of the “w/o Ref” ablation. The original version removed only the final reference addition and retained the reference as a model input. The revised version removes both.

2. Additional Evidence for Source-count Extrapolation

ScaleComposer is trained on two-source configurations, and its reference field is built from surrogates trained on single-source data. All test configurations in both datasets have three or more sources and therefore lie beyond the training range. We provide additional evidence for this extrapolation below.

2.1 Expanded measurements evaluation with two additional six-fan test cases

We have added two six-fan test configurations, expanding the measurement test set from three to five configurations in total. All three six-fan cases share the same spatial layout but differ in their fan rotational speeds (details in Table 3). This allow us to assess whether the observed performance advantage persists across different operating conditions at the highest tested source count, rather than relying on a single speed configuration. The updated evaluation is described in Section 5.1.

2.2 Consistent MAE reductions across configurations and heights in measurements

We now report performance by test configuration and measurement height in Table 2. Compared with Ref+GNN, the strongest baseline in terms of measurement MAE, ScaleComposer reduces mean MAE in every configuration (3%-13%) and at every height (2%-11%). At the same time, its MAE ranges from 0.090 to 0.262 m/s across the three six-fan configurations, despite their identical source count and spatial layout. This variation suggests that prediction difficulty also depends on operating conditions. The highest MAE occurs when all six fans run at full speed, likely because of higher velocities and stronger jet interactions. This case also shows the smallest gain over Ref+GNN (3%). We discuss this further in Section 5.2 of the revised manuscript.

2.3 Full framework outperforms its ablation variants on both datasets.

Tables 9 and 10 show that ScaleComposer achieves the best of all evaluation metrics among the evaluated ablation variants on both datasets. On the numerical benchmark, the representation, attention, and scale-conditioning variants yield similar MAE and RMSE, but lower PCC (0.849 to 0.852, against 0.871). This supports the contribution of the full framework design in both measurement and numerical evaluation. Detailed analyses of the individual components are provided in Response 3.1.

2.4 Performance by source count on both datasets

Figure 3 of the revised manuscript compares ScaleComposer with POD-RBF, Ref+GNN, Ref+GNOT, and Ref+FNO by active-source count. On measurements, ScaleComposer always achieves the lowest MAE. On the numerical benchmark, POD-RBF is more accurate at three sources, the count closest to the training range. ScaleComposer is more accurate as the source count moves further beyond the training range. It also remains below the reference field at every source count on both datasets. In contrast, Ref+FNO and POD-RBF exceed the reference at the highest source counts.

AuthorsOpenReview ↗

General Response to All Reviewers (Part 2/3)

3. Additional Ablations and Baselines

We agree that more comprehensive comparisons are needed. We have added a ROM baseline (see implementation in Table 5). We have updated the results in Table 1 and 7 for the measurement and numerical datasets, respectively. Regarding the ablation study, we have expanded the set of ablation variants, organized them into five categories, and added Table 8 to provide their construction details. The Response 3.1 below discusses the newly added ablations, and their corresponding results are also reported in Table 9 for the measurement dataset and Table 10 for the numerical dataset.

3.1 Additional variants of the two core components

3.1.1 Representation and pooling. To isolate the effects of representation and aggregation, we replaced the proposed two multi-head attention modules with three different representation methods in the table below.

On measurements, ScaleComposer achieves an MAE of 0.180 m/s, compared with 0.189, 0.190, and 0.205 m/s for SourcePool, SourcePairPool, and SourceOnly, respectively. In particular, adding pair representations to source attention reduces MAE by 12.2% relative to SourceOnly. This supports the value of explicitly representing pairwise interactions rather than relying solely on individual source features. Meanwhile, adding pair representations to mean pooling provides no clear improvement, suggesting that simply aggregating additional pair features is insufficient without adaptive weighting.

On the numerical benchmark, the variants exhibit similar MAE (0.033–0.034) and RMSE (0.048–0.049), but ScaleComposer achieves a higher PCC (0.871 vs. 0.849–0.851, a relative increase of 2.35–2.59%). This indicates better agreement with the ground truth in spatial variation, even when the overall pointwise errors are comparable.

Together, these results suggest that explicitly modeling pairwise relationships and adaptively weighting their contributions helps preserve spatial flow patterns, particularly when extrapolating to larger source combinations.

Tested DomainVariantsMAE ↓RMSE ↓PCC ↑
MeasurementSourcePool0.189 ± 0.0040.242 ± 0.0050.784 ± 0.015
SourcePairPool0.190 ± 0.0040.244 ± 0.0020.783 ± 0.010
SourceOnly0.205 ± 0.0250.265 ± 0.0330.745 ± 0.050
Ours0.180 ± 0.0090.234 ± 0.0080.798 ± 0.012
Numerical benchmarkSourcePool0.034 ± 0.0000.049 ± 0.0000.849 ± 0.002
SourcePairPool0.034 ± 0.0000.049 ± 0.0000.851 ± 0.002
SourceOnly0.034 ± 0.0010.049 ± 0.0000.849 ± 0.003
Ours0.033 ± 0.0010.048 ± 0.0010.871 ± 0.003

3.1.2 Attention mechanism or normalization. We replaced the sigmoid gating and null normalization with softmax over valid items, separately for sources and pairs. Standard softmax worsens the measurement MAE from 0.180 to 0.188 m/s and lowers the numerical PCC from 0.871 to 0.851. The reason is that softmax forces the pair weights to sum to one. The single pair seen in training always receives weight one, whereas at test time up to fifteen pairs must share that total, which dilutes each pair’s contribution in a way not seen during training. The sigmoid rule with a null item keeps the single-pair output dependent on its score and lets the total weight grow with the number of relevant items.

Tested DomainVariantsMAE ↓RMSE ↓PCC ↑
MeasurementStandard softmax0.188 ± 0.0080.241 ± 0.0060.793 ± 0.006
Ours0.180 ± 0.0090.234 ± 0.0080.798 ± 0.012
Numerical benchmarkStandard softmax0.035 ± 0.0000.049 ± 0.0000.852 ± 0.003
Ours0.033 ± 0.0010.048 ± 0.0010.871 ± 0.003

3.1.3 Residual reconstruction. We limit the corrections to the configuration-dependent DC term (DC-only), and add the lowest band B1B_1 (Low-frequency), and pool the DC term and B1B_1 on top of them (Low-frequency, pooling). None of the variants reaches the proposed model in terms of MAE. They also correct the magnitude but not the spatial pattern, as their PCC on measurements fall below the reference field 0.785, whereas the proposed model achieves 0.798.

Tested DomainVariantsMAE ↓RMSE ↓PCC ↑
MeasurementDC-only0.216 ± 0.0110.268 ± 0.0090.774 ± 0.009
Low-frequency0.197 ± 0.0190.256 ± 0.0220.741 ± 0.058
Low-frequency, pooling0.189 ± 0.0050.245 ± 0.0080.772 ± 0.024
Ours0.180 ± 0.0090.234 ± 0.0080.798 ± 0.012
Numerical benchmarkDC-only0.036 ± 0.0000.052 ± 0.0000.843 ± 0.000
Low-frequency0.034 ± 0.0000.049 ± 0.0000.849 ± 0.001
Low-frequency, pooling0.035 ± 0.0000.050 ± 0.0000.846 ± 0.001
Ours0.033 ± 0.0010.048 ± 0.0010.871 ± 0.003
AuthorsOpenReview ↗

General Response to All Reviewers (Part 3/3)

3.2 Sensitivity to the reference field construction

As reviewers observed, reference field quality is one of the important factors on the full framework performance. We therefore compare the full framework with reference field, ScaleComposer without reference field, and ScaleComposer with additive reference field. The results are as below.

Table: Ablation of reference field on measurements

VariantsMAE (m/s) ↓RMSE (m/s) ↓PCC ↑
Reference field (Ref)0.2250.2760.785
ScaleComposer w/o Ref0.307 ± 0.0180.450 ± 0.0260.284 ± 0.122
ScaleComposer w/ Additive Ref0.474 ± 0.0470.586 ± 0.0480.701 ± 0.021
Ours0.180 ± 0.0090.234 ± 0.0080.798 ± 0.012

Table: Ablation of reference field on numerical experiments

VariantsMAE ↓RMSE ↓PCC ↑
Reference field (Ref)0.0400.0560.844
ScaleComposer w/o Ref0.046 ± 0.0000.064 ± 0.0010.703 ± 0.003
ScaleComposer w/ Additive Ref0.070 ± 0.0000.095 ± 0.0010.839 ± 0.003
Ours0.033 ± 0.0010.048 ± 0.0010.871 ± 0.003

3.3 Classical reduced-order modeling (ROM) baseline

In addition to previous baselines, we added POD-RBF in the ROM category as one of the traditional CFD surrogate examples.

POD-RBF is competitive on the numerical benchmark, matching ScaleComposer’s aggregate MAE at the reported precision (Table 7). However, the source-count breakdown in Figure 3 of the revised manuscript reveals a clearer distinction: POD-RBF performs better at three sources, whereas ScaleComposer achieves lower MAE at four, five, and six sources, with a larger advantage at five and six sources. At these two higher source counts, POD-RBF also has higher MAE than the reference field, while ScaleComposer improves upon it. It shows that POD-RBF is more accurate when the test set is closer to the training range.

On measurements, ScaleComposer achieves lower MAE than POD-RBF at every evaluated source count (three, four, and six), with overall MAE and RMSE reductions of 53.5% and 54.6%, respectively (Table 1). These results support ScaleComposer’s effectiveness at the larger tested source counts, alongside its substantially stronger performance on the measurement benchmark.

In Table 5, we further present the computational efficiency of this baseline. POD-RBF achieves substantially lower inference latency, owing to its offline–online decomposition. After fitting a reduced POD basis and RBF coefficient mappings, inference requires only coefficient estimation and reconstruction within a fixed low-dimensional subspace, without iterative neural-network evaluation. However, this efficiency comes with constraints on its representation and extrapolation capacity. The POD basis is derived from available training snapshots, while RBF interpolation must estimate coefficients for unseen configurations with higher source counts, where interacting jets may introduce spatial structures insufficiently represented in the training data. This interpretation is consistent with the source-count breakdown in Figure 1 of paper.

4. Additional Experiments

4.1 Performance break-down to assess robustness

The measured test set now has five configurations, including three six-fan cases. We updated with breakdown performance by configurations and by heights in Table 2. We compared the proposed ScaleComposer with the strongest baseline Ref+GNN, ours shows lower mean MAE in all five configurations and at all four heights. Please find results and detailed analysis in Section 5.2.

4.2 Sensitivity analysis of the frequency-band design

We compare the original partition with variants in which we scaled the internal band boundaries by 0.8 and 1.2. The measurement MAE of all band boundaries remains comparatively stable within 0.179-0.181 m/s. On numerical, the original partition is the best of the three (Table 12), but performance becomes more sensitive to shifting the band boundaries toward lower frequencies. Please find full results and analysis in Appendix A.5.2.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.