acceptodds
Under review as a conference paper at ICLR 2027

FluidWeave: Reconstructing Physical Fields from Sparse Sensors via Gaussian Primitives

Abstract

Reconstructing continuous physical fields from sparse surface-mounted sensors is central to scientific monitoring, engineering design, and digital-twin systems. Many neural reconstructors predict queried field values through latent features that do not themselves define a field over the target domain. To make this spatial structure explicit, we introduce FLUIDWEAVE, an end-to-end framework built around a shared spatial state that couples a continuously queryable field proposal with its residual correction. Inspired by 3D Gaussian Splatting, we instantiate this state as K anisotropic Gaussian primitives predicted from the unordered sensor set, whose centers, covariances, and amplitudes encode location, support, and field value, respectively. Their partition-of-unity aggregation forms the normalized proposal, while the full primitive state conditions the correction decoder. To analyze this proposal-correction mechanism, we derive a coverage-amplitude error bound, decompose local sensor sensitivity across the two pathways, and establish an exact identity for the residual's error reduction. Across four complementary benchmarks, from 2D cylinder flow and AirfRANS to 3D FlowBench LDC-3D and PhySense-Car, FLUIDWEAVE outperforms several strong reconstruction baselines.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Scientific machine learning, physical-field reconstruction, sparse surface-mounted sensors, inverse problems, Gaussian primitives, continuous field representations

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_3bpxOpenReview ↗
Rating
3
Confidence
4(+1)
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes FLUIDWEAVE, a framework for reconstructing continuous physical fields from sparse sensor observations. It predicts per-instance anisotropic Gaussian primitives that jointly define a normalized field proposal and condition a residual decoder. The authors provide theoretical analyses of proposal error, sensor sensitivity, and residual correction, and demonstrate strong reconstruction performance across four 2D and 3D flow benchmarks.

Critical strengths

S1. Meaningful structural inductive bias. The shared Gaussian representation provides an explicit, continuously queryable field proposal while conditioning the residual decoder. This is a useful architectural contribution that introduces spatial structure into sparse-sensor reconstruction, rather than relying entirely on implicit latent representations.

S2. Strong empirical performance across diverse benchmarks. The method achieves the lowest mean reconstruction error among the evaluated approaches on Cylinder, AirfRANS, FlowBench LDC-3D, and PhySense-Car. These benchmarks cover regular grids, irregular meshes, 3D volumes, and unstructured surfaces, demonstrating the applicability of the proposed framework across different spatial representations.

S3. Comprehensive architectural ablations. The paper includes matched direct-decoder, generic-token, kernel-replacement, and pathway ablations. The substantial degradation when removing state attention or the shared spatial representation supports the effectiveness of the proposed architecture beyond simply increasing model capacity.

S4. Careful experimental protocols and diagnostic analyses. Multiple training seeds, consistent sensor layouts, channel-balanced metrics, sensor perturbation studies, and detailed implementation settings strengthen the reliability of the results. The additional Jacobian and correction analyses provide useful insights into model behavior.

Critical weaknesses

W1. Unclear conditioning parity and sensor dependence. On AirfRANS, FLUIDWEAVE uses operating metadata and rich geometric query features, including signed distance, surface normals, and flow-frame coordinates. However, it is not sufficiently clear whether all baselines receive equivalent conditioning information. Moreover, sensor-removal or sensor-shuffling experiments with geometry and metadata fixed are missing for AirfRANS, making it difficult to fully isolate the contribution of sparse sensor observations.

W2. Insufficient isolation of adaptive Gaussian geometry. Although the generic-token and kernel controls are valuable, they do not directly establish whether predicting per-instance Gaussian centers and covariances is necessary. A matched comparison against fixed or globally learned Gaussian geometry with sensor-conditioned amplitudes is missing. Comparisons with closely related continuous reconstruction architectures, such as ConvCNP or conditional INRs, would also better establish the specific contribution of the proposed representation.

W3. Limited explanatory connection between the theoretical analysis and empirical improvements. The theoretical propositions are mathematically valid but largely follow standard approximation bounds and algebraic decompositions. Empirically, the coverage-error Pearson correlation is only 0.1173, while the residual pathway dominates local sensor sensitivity. These results do not invalidate the approach, but leave the precise role of the explicit Gaussian proposal in the final prediction insufficiently characterized.

W4. Unexplained primitive-count and proposal-quality trade-offs. In the Cylinder Surface-8 ablation, K=64 achieves a lower full-model error (0.0282) than the selected K=384 configuration (0.0358), despite producing a less accurate standalone proposal. This suggests that improving proposal accuracy does not necessarily improve final reconstruction. The rationale for selecting the larger configuration and the associated accuracy-efficiency trade-offs are not sufficiently explained.

W5. Some generalization and physical-fidelity results require further qualification. On AirfRANS Re-shift, FLUIDWEAVE (0.3129 ± 0.0102) performs only marginally better than FLRNet (0.3142 ± 0.0018), making the strength of this advantage uncertain. On LDC-3D, the w-velocity channel retains approximately 0.65 relative L2 error across sensor counts. Although overall and channel-balanced results are strong, the extent to which the reconstructed fields preserve important physical structures remains unclear.

Questions for authors

Q1. Regarding W1: Could the authors clarify whether all AirfRANS baselines receive identical geometric features and operating metadata? Can they provide sensor-removal or sensor-shuffling experiments while keeping these conditions fixed? Such results would help establish that the improvements originate from better utilization of sparse observations.

Q2. Regarding W2: Could the authors compare FLUIDWEAVE against a matched variant with fixed or globally learned Gaussian centers/covariances and per-instance predicted amplitudes? A comparison with ConvCNP or conditional INR under matched conditioning would further clarify the benefit of adaptive Gaussian geometry.

Q3. Regarding W3: Could the authors provide stronger empirical evidence connecting the proposed theoretical quantities to reconstruction quality, particularly the relationship between coverage, proposal error, and final prediction error? How should the dominant residual sensitivity be interpreted in terms of the functional contribution of the Gaussian proposal?

Q4. Regarding W4: Why was K=384 selected for Cylinder Surface-8 when K=64 achieves better test accuracy? Could the authors report the corresponding validation performance and inference costs, and explain why improving standalone proposal accuracy does not consistently improve the final model?

Q5. Regarding W5: Could the authors quantify the statistical reliability of the improvements under AirfRANS distribution shifts using paired comparisons? For LDC-3D, could they provide additional velocity-specific or physical-consistency diagnostics and explain the persistently high w-channel reconstruction error?

6 comments
AuthorsOpenReview ↗

Respone W1/Q1 to Equal Conditioning and Sensor-Dependent Reconstruction

Thank you for these questions. All AirfRANS baselines receive the same conditioning information. Our additional sensor-shuffling tests show that FluidWeave’s reconstruction depends on the correspondence between pressure readings and sensor locations, even when geometry and operating metadata remain fixed.

Conditioning parity. All compared methods, including the newly added ConvCNP and PhySense, receive the same sensor observations, operating metadata, and geometric query features, including signed distance, surface normals, and flow-frame coordinates. FluidWeave therefore does not have exclusive access to these conditioning inputs.

Sensor shuffling on AirfRANS Full-8 (Table 5). We use the same three trained FluidWeave models and test set as the full-model reference in Table 5. We permute the eight pressure readings among their fixed sensor locations, while keeping geometry, operating metadata, query points, and trained weights unchanged.

Sensor readingsMean MSERelative increase
Original readings (Table 5, full model)0.2367—
Shuffled: 8 permutations0.288621.9%
Shuffled: 16 permutations0.283319.7%

Results average 200 test cases and three trained models.

These stress tests show that disrupting the reading–location correspondence can substantially degrade reconstruction despite unchanged geometry and operating metadata.

Complementary metadata-only comparison. We also evaluate a separately trained metadata-only baseline on LDC-3D to assess the additional predictive value of sensor observations after retraining. Its protocol and results are detailed in our response to Reviewer 1Xsr.

Manuscript changes (highlighted in blue in the revised PDF). Appendix J.1 (p. 29) explicitly states that FluidWeave and all compared AirfRANS baselines, including ConvCNP and PhySense, receive the same sensor observations, operating metadata, and geometric query features. Section 5.3 (p. 10) summarizes reading-location dependence and the complementary retrained LDC controls. Appendix J.4 (pp. 30-31) specifies the fixed inputs and permutation protocol, with both shuffling budgets in Table 29 (p. 31) and the no-reading controls in Table 28 (p. 31).

AuthorsOpenReview ↗

Response to W2/Q2: Gaussian geometry and New baselines

Thank you for these questions. We have completed the geometry control and ConvCNP comparisons across all configurations in Tables 1–3.

Globally learned geometry. We share Gaussian centers and covariances across cases while retaining sensor-conditioned amplitudes and weights. The primitive count, decoder, losses, and training budget are unchanged.

VariantFull-8 test MSE
Globally learned geometry0.280371 ± 0.034595
FLUIDWEAVE0.236730 ± 0.026067

Mean ± SE over three runs.

ConvCNP. We evaluate all ten configurations with five runs each, using the corresponding splits, sensor layouts, and metrics, with geometry and operating-condition inputs where applicable.

Cylinder: relative L2L^2

MethodSurface-4Surface-8Surface-16
ConvCNP0.0733 ± 0.01500.1249 ± 0.00860.0732 ± 0.0015
FLUIDWEAVE0.0642 ± 0.00180.0358 ± 0.00330.0160 ± 0.0003

AirfRANS: normalized-field MSE

MethodFull-8AoA-shift-8Re-shift-8
ConvCNP0.2822 ± 0.02060.6060 ± 0.06590.3298 ± 0.0209
FLUIDWEAVE0.2526 ± 0.01860.5280 ± 0.04520.3129 ± 0.0102

LDC-3D: global and channel-macro relative L2L^2

SensorsConvCNP: GlobalFLUIDWEAVE: GlobalConvCNP: MacroFLUIDWEAVE: Macro
50.0681 ± 0.00160.0303 ± 0.00030.4027 ± 0.00510.2246 ± 0.0032
100.0632 ± 0.00160.0284 ± 0.00030.3969 ± 0.00510.2229 ± 0.0035
200.0603 ± 0.00130.0281 ± 0.00040.3930 ± 0.00410.2247 ± 0.0034
400.0590 ± 0.00170.0273 ± 0.00040.3919 ± 0.00530.2225 ± 0.0037

Mean ± SE over five runs.

FLUIDWEAVE achieves lower mean error in all ten configurations, including both AirfRANS shifts and both LDC-3D metrics. These results strengthen the comparison with continuous functional reconstruction across sensor densities and spatial discretizations.

Manuscript changes (highlighted in blue in the revised PDF). ConvCNP and PhySense appear in all ten configurations of Tables 1-3 (p. 7), with conditioning and protocols in Appendix J.1 (p. 29). Section 5.3 (pp. 9-10) distinguishes generic-token, radial-response, and shared-geometry controls. Appendix J.5 (p. 31) specifies the globally learned centers/covariances, sensor-conditioned amplitudes/weights, retained design, parameter counts, and ten-epoch training budget; Table 30 (p. 31) reports the three-run AirfRANS Full-8 geometry comparison (0.280371 vs. 0.236730).

AuthorsOpenReview ↗

Reponse to W3/Q3: Theory and Evidence

Thanks for these questions. We have strengthened this connection through a coverage-stratified analysis of the existing Table 5 evaluations, linking proposal error, correction gain, and reconstruction improvements on the same cases.

Coverage, error, and reconstruction gains. We use the same checkpoints and all 200 AirfRANS Full-8 test cases as Table 5, evaluating all native query points across three runs. Cases are grouped by their mean coverage radius across the full models. The table reports normalized-field MSE averaged over cases and runs.

Prediction / configurationAll 200 casesLow radius: 67Medium radius: 67High radius: 66
Internal proposal SS (full model)0.99050.68810.93001.3589
Full model0.23670.15360.21240.3459
No additive proposal0.26820.16490.23770.4040
No state features0.29920.19540.26990.4343
No state attention0.34460.24390.30970.4824

The four configuration rows reproduce the Table 5 means in the All 200 cases column; weighting the unrounded group means by case counts recovers these results. Internal SS is taken from the jointly trained full model, distinct from the separately trained Proposal only ablation. No additive proposal retains state-conditioned correction.

Larger-radius groups have higher proposal and final errors, together with larger mean absolute penalties for removing the additive proposal or state conditioning. The full model achieves lower mean MSE than each control in every group. The correction identity quantifies the intermediate improvement: the mean per-case fraction of proposal MSE removed by correction is 75.5%, 72.0%, and 63.4%, respectively.

Across cases, proposal and final MSE have Pearson correlations of 0.880, 0.915, and 0.884 in the three runs. These measure casewise proposal-to-final error association, whereas the original 0.1173 measures pointwise coverage-to-proposal error association. Our bound separates coverage and amplitude discrepancy; this joint analysis connects the geometric diagnostic, the resulting proposal error, and its correction to final reconstruction quality and the observed ablation gains.

Interpreting residual sensitivity. The residual Jacobian includes sensor dependence propagated through the shared Gaussian state, as well as through other conditioning inputs. Its dominance measures the local response to sensor perturbations, rather than the accuracy contribution of removing the state. The controls above show that retaining the additive proposal and state conditioning benefits the correction-based prediction. Thus, the residual performs substantial correction while benefiting from the explicit spatial state.

Manuscript changes (highlighted in blue in the revised PDF). Section 5.3 (p. 10) connects coverage, internal-proposal error, correction gains, and final accuracy, distinguishing the internal proposal from the separately trained proposal-only control. Appendix J.6 (pp. 31-33) specifies the common cases, grouping, averaging, and run-wise correlations; Table 31 (p. 32) reports all grouped controls and correction gains. Section 5.3 (p. 10) explains state-mediated residual sensitivity, with the complete noise/sensitivity plot in Figure 6 (p. 32) and ratios in Table 32 (p. 33).

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

This paper proposes FLUIDWEAVE for reconstructing continuous physical fields from sparse surface-mounted sensors. The method predicts a set of anisotropic Gaussian primitives from an unordered sensor set, aggregates them through a partition-of-unity scheme to form a continuously queryable field proposal, and feeds the complete primitive state into a residual decoder for correction. The authors further analyze the proposal in terms of coverage and amplitude errors, sensor sensitivity, and the error reduction induced by residual correction. On four two-dimensional and three-dimensional benchmarks, namely Cylinder, AirfRANS, FlowBench LDC-3D, and PhySense-Car, FLUIDWEAVE achieves competitive reconstruction performance relative to the evaluated methods. Overall, the paper presents a clearly structured and reasonably novel framework for spatially structured field reconstruction.

Critical strengths
  1. The experiments cover regular grids, irregular meshes, two-dimensional fields, three-dimensional volumetric fields, and unstructured surfaces, providing a broad evaluation of the proposed framework across different spatial representations.

  2. The authors report multiple independent training runs, ablation studies, sensor-noise experiments, and per-channel diagnostics, resulting in a relatively comprehensive empirical evaluation.

  3. The Gaussian primitive state offers potential for visualization and intervention, providing an interface for analyzing how information from sparse sensors is allocated across the target spatial domain.

Critical weaknesses
  1. The paper does not report results for PhySense on the Cylinder, AirfRANS, or FlowBench LDC-3D benchmarks. The absence of this closely related baseline makes it difficult to fully assess the relative advantage of FLUIDWEAVE. The authors should evaluate the original PhySense method, or provide a faithful reimplementation, on the first three datasets using the same data splits, sensor layouts, sensor counts, and evaluation protocols.

  2. The paper does not evaluate the PhySense sea-temperature benchmark, which constitutes a challenging test setting. The authors should conduct experiments on this benchmark following the official protocol and report results for FLUIDWEAVE and the main competing methods. This would provide stronger evidence for the method’s effectiveness and generalization ability.

  3. The paper interprets the Gaussian primitives as a physically meaningful spatial state or field proposal, but the primitive parameters are not directly supervised by physical targets, and the training objective only constrains the final sum S+RS+R. The current experiments provide stronger support for the usefulness of a structured spatial intermediate representation than for a physically meaningful proposal–correction decomposition.

Questions for authors

See Weaknesses.

3 comments
AuthorsOpenReview ↗

Response to W1: PhySense Comparisons

Thank you for this suggestion. We have completed PhySense comparisons across all ten configurations of Cylinder, AirfRANS, and FlowBench LDC-3D.

We adapt the official PhySense implementation, retaining its reconstruction backbone and flow-matching training objective, and select checkpoints using validation data. Comparisons use the same data splits, fixed sensor layouts, sensor counts, and evaluation protocols as FLUIDWEAVE.

The available conditioning information is also provided to PhySense: AirfRANS includes sensor/query geometry features, geometry tokens, and case metadata; LDC-3D includes query geometry (SDF) and Reynolds number, alongside sensor measurements and coordinates. Thus, the comparisons do not withhold geometry or operating-condition information from PhySense.

All results below are mean ± SE over five independent training runs per setting; lower is better.

Cylinder Physical-space relative L2L^2.

SettingPhySenseFLUIDWEAVE
Surface-40.0857 ± 0.01220.0642 ± 0.0018
Surface-80.0870 ± 0.01200.0358 ± 0.0033
Surface-160.0731 ± 0.00890.0160 ± 0.0003

AirfRANS Aggregate normalized-field MSE.

SettingPhySenseFLUIDWEAVE
Full-80.2784 ± 0.01740.2526 ± 0.0186
AoA-shift-80.5788 ± 0.03860.5280 ± 0.0452
Re-shift-80.3583 ± 0.01220.3129 ± 0.0102

FlowBench LDC-3D Global and channel-balanced macro relative L2L^2.

SensorsPhySense: globalFLUIDWEAVE: globalPhySense: macroFLUIDWEAVE: macro
50.03926 ± 0.001130.03032 ± 0.000350.27070 ± 0.005300.22464 ± 0.00322
100.04750 ± 0.008290.02840 ± 0.000320.35738 ± 0.087720.22287 ± 0.00348
200.03638 ± 0.000760.02814 ± 0.000350.26018 ± 0.003990.22465 ± 0.00341
400.03402 ± 0.001600.02728 ± 0.000410.25651 ± 0.012170.22249 ± 0.00373

FLUIDWEAVE achieves lower mean error in all ten configurations, including both AirfRANS distribution shifts and both LDC-3D metrics.

Manuscript changes (highlighted in blue in the revised PDF). PhySense results are included in Tables 1-3 (p. 7). Appendix J.1 (p. 29) specifies its released backbones, flow-matching objective, conditioning, fixed layouts, and validation-based checkpoint selection. Table 24 (p. 29) and Table 25 (p. 29) report the full five-run LDC global and channel-macro summaries.

AuthorsOpenReview ↗

Response to W2: Sea-Temperature Benchmark

Thank you for this suggestion. We have added the PhySense sea-temperature benchmark, using its officially released training/test data, ocean mask, time inputs, and random ocean-sensor sampling protocol. The release contains 9,843 training fields and 366 test fields on a 171x360 grid.

We train FLUIDWEAVE with 10–100 randomly sampled sensors and evaluate the same model at all four sensor counts. FLUIDWEAVE and the released PhySense BASE checkpoint receive identical sensor locations, observations, and available auxiliary information, including the ocean mask and time inputs.

Method15 sensors25 sensors50 sensors100 sensors
PhySense, released checkpoint0.0521680.0469480.0452040.044123
FLUIDWEAVE0.0457720.0453600.0450500.044186

Mean per-field relative �0��0� over ocean cells, averaged over five shared random-layout schedules covering all 366 test fields. Evaluations use one trained FLUIDWEAVE model and one released PhySense checkpoint; PhySense errors are additionally averaged over three inference-noise draws.

FLUIDWEAVE reduces mean error by 12.26% at 15 sensors and 3.38% at 25 sensors, with similar mean errors at 50 and 100 sensors. It uses 0.220M parameters versus 19.060M, approximately 1.15% of the PhySense reconstructor’s parameter count.

These results extend our evaluation to sea-temperature reconstruction with complex land–sea boundaries, with the clearest gains under sparse sensing.

Manuscript changes (highlighted in blue in the revised PDF). The Abstract (p. 1), Introduction (p. 2), and Section 6 (p. 10) use the five-benchmark scope. Section 5.1 (p. 8) reports the Sea gains and parameter counts; Table 5 (p. 9) separates published results from local evaluation. Appendix J.2 (p. 30) specifies the data, shared layouts and auxiliary inputs, one-model evaluation, and averaging over layouts/noise draws; Table 26 (p. 30) includes the complete published comparison.

AuthorsOpenReview ↗

Response to W3: Spatial Inductive Bias

Thank you for raising this distinction. Our claim is a structured spatial intermediate representation that provides an inductive bias, as motivated in the Introduction, rather than a decomposition into distinct physical mechanisms.

The Gaussian centers and covariances explicitly define spatial location and support, while the amplitudes parameterize a queryable field proposal. These parameters are learned jointly through reconstruction supervision; they are not intended to identify individual physical entities.

The contribution lies in using the same spatial state both to construct the additive proposal and to condition its correction. The additive-proposal and state-conditioning controls in Table 5 evaluate these computational roles. Thus, the structured-representation interpretation highlighted by the reviewer is consistent with our intended contribution.

Manuscript changes (highlighted in blue in the revised PDF). Section 3.2 (p. 3) identifies the learned primitive state and its field-valued proposal, explaining their coupled evaluation/correction roles. Section 5.3 (p. 9) and Table 6 (p. 10) provide the additive-path and state-conditioning evidence; the original Table 5 is now Table 6.

Reviewer_1XsrOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper presents FluidWeave, a method for reconstructing physical fields from sparse sensor measurements. It uses a set of anisotropic Gaussian primitives to represent a continuous spatial field, together with a residual decoder based on cross-attention to improve the prediction. The authors also provide some theoretical analysis of the proposal and residual components. Experiments on four benchmarks, covering both 2D and 3D settings, show that FluidWeave outperforms the compared baselines in most settings.

Critical strengths
  • The method is flexible and works on different types of grids and meshes, including both 2D and 3D data.

  • The experiments are extensive, with multiple benchmarks, ablations, and noise tests.

  • The results show consistent improvements over the compared baselines, especially on LDC-3D and PhySense-Car.

Critical weaknesses
  1. It is not clear how much the Gaussian proposal actually contributes. The proposal alone has much higher error than the full model, and the Jacobian analysis shows that most sensor sensitivity comes from the residual decoder. Also, the generic-token and no-additive-proposal ablations perform quite close to the full model, considering the reported standard errors. It seems possible that the main improvement comes from the attention decoder rather than the explicit Gaussian representation?

  2. The theoretical analysis is rather limited. Most propositions follow directly from basic mathematical properties, such as the triangle inequality, chain rule, and expansion of squared norms. They do not really explain why the proposed architecture should work better, or provide much guidance for the design.

  3. Several relevant baselines, such as Transolver, GINO, and conditional INR methods, are missing. On PhySense-Car, even the locally evaluated baselines are much better than previously published numbers, which raises concerns about the evaluation protocol. For LDC-3D, removing sensor values barely affects the velocity predictions, suggesting that the model may rely mainly on metadata.

Questions for authors
  1. Why was K=384 selected for the main AirfRANS results when K=64 gives better performance in Table 12? Also, could the authors explain why a better proposal does not lead to a better final prediction?

  2. Could the authors clarify the evaluation protocol on PhySense-Car and the role of metadata on LDC-3D? In particular, a metadata-only baseline would be helpful to understand how much information is actually reconstructed from the sparse sensors.

5 comments
AuthorsOpenReview ↗

Response to Critical Weaknesses 1(W1):Gaussian state contribution

FluidWeave’s Gaussian state jointly generates the proposal and conditions the correction decoder. Proposal-only accuracy therefore does not capture its full role. Likewise, the residual Jacobian includes state-mediated sensor dependence, so its dominance cannot isolate the attention decoder’s contribution.

The following AirfRANS Full-8 controls examine the state’s role:

Control or interventionResultSource
Full modelMSE 0.2367 ± 0.0261Table 5
Remove proposal value and basis mass from decoder inputs; retain attention and final additionMSE 0.2992 ± 0.0099Table 5
Remove final proposal addition; retain state-conditioned correctionMSE 0.2682 ± 0.0086Table 5
Swap primitive state across cases; retain recipient context, geometry and metadataMean global relative L² increases by 0.0276, from 0.3151Appendix H.3

With attention and the final proposal addition retained, removing the explicit decoder inputs gives a higher mean MSE, providing evidence for the value of proposal-derived information alongside attention. Separately, swapping the primitive state across cases, with recipient context, geometry and metadata unchanged, increases the mean error (averaged over 200 test cases, three models and ten donor permutations), showing that the complete prediction depends on the case-specific state.

Generic-token routing (0.2627; Table 5) and Wendland C2 (0.2745; Appendix Table 15) also yield higher mean MSE than the full model. Across these training ablations, the full model attains lower mean errors, but the three-run estimates do not establish statistically reliable superiority.

Manuscript changes (highlighted in blue in the revised PDF). Section 5.3 (p. 9) explains the retained and removed state pathways, the same-objective direct-decoder comparison, and the complementary state-swap result. The original Table 5 is now Table 6 (p. 10); intervention details remain in Appendix H.3 (pp. 23-24). Section 5.3 (pp. 9-10) discusses generic-token and radial-response controls; the original Appendix Table 15 is now Table 16 (p. 26). State-mediated residual sensitivity is explained in Section 5.3 (p. 10).

AuthorsOpenReview ↗

Response to W2: The theoretical analysis

The analysis gives testable criteria for proposal quality, correction, and sensor sensitivity. Their practical implications and supporting evidence are summarized below.

1. Proposal quality: assess coverage and predicted values together.

The proposal bound separates spatial coverage from amplitude error, indicating that geometry and predicted values should be assessed jointly.

New analysis: we evaluate the same Table 5 full-model checkpoints on all 200 AirfRANS Full-8 cases and all native queries. Cases are grouped by their three-model mean coverage radius: the weighted query-to-primitive-center distance. Larger radii mean more distant contributing centers.

Coverage-radius groupCasesMean radiusProposal MSEFinal MSE
Low670.14370.68810.1536
Medium670.15690.93000.2124
High660.18671.35890.3459
All cases2000.16230.99050.2367

MSE is averaged over cases and three runs in normalized field units. Proposal denotes internal SS of the full model, not the separately trained Proposal only ablation. The all-case final MSE matches Table 5.

Larger-radius groups have higher mean proposal and final errors, linking the spatial diagnostic to both stages of reconstruction.

2. Correction: assess its direction and size, not size alone.

The correction identity relates error reduction to alignment with the proposal error and relative correction magnitude. Our correction decoder receives proposal values and primitive information. The existing Table 5 controls below test whether this state interface and the additive proposal help:

ConfigurationMSE, mean ± SE over three runs
Full model0.2367 ± 0.0261
No additive proposal; state conditioning retained0.2682 ± 0.0086
No state features: proposal value and basis mass0.2992 ± 0.0099
No state attention0.3446 ± 0.0083

Same 200 AirfRANS Full-8 cases and normalized-field MSE; three training runs per configuration. SE is computed across runs.

The full model has the lowest mean error among these controls, supporting both the additive and conditioning roles of the state.

3. Sensitivity: predict how sensor noise changes the output.

The sensitivity analysis includes both pathways and their interaction. The existing small-noise experiment compares predicted and measured output changes after adding Gaussian sensor noise, with geometry and operating conditions fixed.

Sensor-noise levelMeasured output change / predicted output change
0.5%1.045
1%1.143
2%0.903

A ratio of 1 means agreement. Output change is the mean-squared difference between noisy and clean predictions, not reconstruction error. Ratios compare aggregated means: three models, 30 cases/model, 512 queries/case, and 16 noise draws/case/model. Noise standard deviations are 0.005, 0.01, and 0.02 in normalized sensor units.

The ratios show that the local analysis quantitatively tracks the observed small-noise response.

Manuscript changes (highlighted in blue in the revised PDF). Section 4 (pp. 5-6) states the practical interpretations of the coverage bound, correction identity, and sensitivity decomposition. Section 5.3 (p. 10) connects these quantities to proposal/final errors and correction gains; Section 5.3 (p. 10) reports the small-noise agreement. Appendix J.6 (pp. 31-33) gives the grouping and aggregation protocols, Table 31 (p. 32) the grouped errors and controls, and Table 32 (p. 33) the measured/predicted ratios. The former main-text Figure 4 is now Figure 6 (p. 32) in that appendix.

AuthorsOpenReview ↗

Response to Q1: Primitive Count and Proposal Quality

Thank you for this question. AirfRANS uses K=96K=96, not K=384K=384, as specified in Appendix F.3. Table 12 concerns Cylinder Surface-8. We provide the selection evidence and ablation results for that setting below.

Why K=384K=384? The main Cylinder configuration was selected through a finite joint configuration search, rather than by optimizing KK alone. The K=384K=384 configuration achieved the lowest validation error among the following candidates; K=64K=64 was not included in this initial search.

KKDecoder widthAttention dimensionValidation relative L2L^2
192256640.07408
256320960.06844
3203841280.05101
3844481600.04410

One training run per candidate, evaluated on 200 validation frames. These were jointly varied configurations, also involving learning rate, observation-loss weight, and training-query count.

Fixed-configuration KK-ablation. Table 12 subsequently varies KK while retaining decoder width 448, attention dimension 160, and the remaining settings:

KKFull-model test relative L2L^2Scaffold-only test relative L2L^2
640.0282 ± 0.00080.3471 ± 0.0067
1280.0287 ± 0.00070.2933 ± 0.0113
2560.0298 ± 0.00150.2501 ± 0.0056
3840.0358 ± 0.00330.2356 ± 0.0089

Mean ± SE over five independent training runs on the disjoint 4,750-frame test set, as reported in Table 12.

The fixed-configuration sweep indeed favors smaller primitive counts. This distinguishes selecting a complete configuration within the initial candidate set from assessing KK with the other settings fixed.

Why does a better proposal not guarantee a better final prediction? The Gaussian state has two coupled roles: it constructs SS and conditions RR. Changing KK therefore changes both the proposal and the information available to the correction pathway. Joint training optimizes S+RS+R, so standalone proposal error alone does not determine final reconstruction error. Table 12 illustrates this distinction: increasing KK improves the scaffold-only result but not the full-model result.

Manuscript changes (highlighted in blue in the revised PDF). Appendix J.7 (p. 33) distinguishes AirfRANS K=96 from Cylinder Surface-8. Table 33 (p. 33) documents the finite joint search and single-run, 200-frame validation scores. The original Table 12 is now Table 13 (p. 24); Table 34 (p. 33) reports its five-run test results with inference costs. The accompanying text distinguishes configuration selection from the K-only sweep and explains why proposal-only accuracy does not determine final accuracy.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary for the Area Chair and Reviewers

We thank the Area Chair and reviewers for their constructive feedback. The revision strengthens the evidence for FluidWeave's central contribution and integrates the requested analyses into the manuscript.

Core contribution. For domain-wide reconstruction from sparse surface observations, FluidWeave predicts a per-case Gaussian spatial state that serves two coupled roles: it forms a continuously queryable, normalized field proposal and conditions the residual correction. The architectural contribution is this shared use of an explicit spatial state in both prediction and correction.

Main clarifications and supporting evidence:

  1. Contribution of the shared state. On AirfRANS Full-8, removing the additive proposal while retaining state-conditioned correction increases the three-run mean MSE from 0.2367 to 0.2682; removing proposal-derived correction inputs raises it to 0.2992. These controls support both computational roles. Generic-token, radial-response, and globally learned geometry comparisons further examine the state representation.

  2. Closest baselines and broader evaluation. We completed five-run ConvCNP and PhySense comparisons across all ten Cylinder, AirfRANS, and LDC-3D configurations; FluidWeave achieves lower mean error than both methods in each configuration. The added PhySense-Sea benchmark extends evaluation to ocean-temperature reconstruction.

  3. Information supplied by sensors. Retrained LDC-3D controls without readings have higher mean errors in all velocity components and pressure. AirfRANS shuffling, with geometry and operating metadata fixed, separately shows that the trained model uses reading-location correspondence.

  4. Connection between analysis and accuracy. On the same 200 AirfRANS test cases, we relate coverage to internal-proposal error, correction gains, final error, and state-ablation results. Small-noise measurements also track the sensitivity predictions, connecting the theoretical quantities to observed model behavior.

  5. Physical-field diagnostics and evaluation protocols. Additional velocity, vorticity, channel-wise, and loss-weighting analyses characterize the LDC-3D reconstructions. We clarify conditioning inputs and the PhySense-Car evaluation, and distinguish validation-based configuration selection from the primitive-count accuracy-efficiency study.

The revised main text presents the resulting conclusions, with detailed protocols and results in the appendices. Manuscript changes are highlighted in blue, and each response below ends with its corresponding revised sections, tables, and pages.

Dear reviewers,

The author-reviewer discussion period is ending soon. If you have not done so yet, please read the authors’ rebuttal and reply to it. Tell the authors whether it addresses your concerns, and update your score if your assessment has changed. With the high number of submissions this year, your input matters more than ever.

Thank you for your time and help.

Best regards,


Your Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.