acceptodds
Under review as a conference paper at ICLR 2027

Where Should Observations Enter? Dual-Path Integration for Continuing Sensing with Recurrent Gaussian States

Abstract

Reconstructing dense unsteady flow fields from sparse near-body measurements is an underdetermined inverse problem, as most of the field is never directly observed. A common latent-state design for sparse flow reconstruction compresses the available measurements into a single latent state and decodes every requested coordinate from that state. Although efficient for temporal propagation, this design couples state updating with query readout: when the latent state is the decoder's sole interface to the observations, every query is decoded from the same fused summary. This raises a central design question: should observations enter the recurrent state, the query readout, or both? GaussAssim addresses this question with two observation paths built around a recurrent anisotropic Gaussian field state. The state path assimilates observations into the Gaussian state for current decoding and subsequent propagation, while the query path uses time-indexed observation tokens to correct each readout without altering that state. On the ACDM Inc and periodic FlowBench benchmarks, GaussAssim outperforms the compared learned baselines in mean reconstruction error. Across four evaluation settings, ten-seed factorial comparisons show that each observation path further reduces mean error when the other is active. All eight conditional gains have pointwise 95% bootstrap intervals above zero. Equal-token controls on NURBS show that distinct past observations contribute beyond repeated current frames, while spatial diagnostics associate larger query corrections with higher state-only error. Together, these results demonstrate complementary benefits from recurrent state assimilation and direct query-conditioned observation access in continuing-sensing flow reconstruction.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Sparse flow reconstruction, Continuing sensing, Unsteady flow, Recurrent Gaussian states, Observation assimilation, Query-conditioned correction, Spatiotemporal field reconstruction, Scientific machine learning

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_FjLkOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper reconstructs unsteady flow fields from sparse near-body sensors (16 probes, 4-snapshot window). It asks whether new observations should update a recurrent state, feed the query-time readout directly, or both. GaussAssim does both: (1) State path: a recurrent set of Gaussian primitives that assimilates each snapshot and decodes a base field, and (2) Query path: each query coordinate attends to the raw sensor–lag tokens and adds a correction that is not written back to the state.

The method outperforms the compared baselines on three periodic FlowBench families and on Reynolds-held-out ACDM Inc. The mechanism is supported by three experiments: a ten-seed 2×2 factorial, a parameter-matched shared-summary vs. query-conditioned readout comparison, and temporal-input controls.

Critical strengths

S1. Clear accuracy gains on periodic flows. On the three FlowBench families, the method reduces mean error from 0.1149 (state-only) to 0.0889, and is about 22% below the strongest compared baseline (0.1146, Table 1). The gain holds in every family and in different seeds (Tables 13–15). The margin on ACDM is much smaller.

S2. The query-conditioned readout is cleanly attributed. With matched architecture, training, and parameter count, per-query aggregation beats a shared observation summary by 6–7%. It improves 82/84 geometries and 11/11 Reynolds cases (Table 5, Figure 4). A parameter-matched head without observation tokens also trails the full model (Table 9). This is a clear, reusable design finding.

Minor: the authors also report some results that do not favor their method (Tables 12, 13, 15, 21), which I appreciate.

Critical weaknesses

W1. The title question is only partly answered. (1) The recurrence is too short to test the claim. It spans only four snapshots and re-initializes every window. The key design choice, not writing the correction back into the state, is therefore never tested against the alternative. (2) On ACDM, history adds little. Current-frame encoding nearly matches the full model (0.0201 vs. 0.0195). Replacing the query input with one, wrong-lag, or repeated snapshots changes error by at most 0.0002 (Table 9). (3) The design is close to Attentive Neural Processes. A compressed summary plus a per-query attention path to raw context is the core ANP idea.

The supported conclusion is therefore narrower than claimed: per-query attention on top of a short-window state helps, mainly on periodic flows. To address: (1) a write-back variant; (2) longer or streaming horizons without re-initialization (ACDM has 1,300-step trajectories); (3) a clearer positioning against ANPs. Positive results on (1) and (2) would raise the score.

W2. The benefit of the query correction is narrower than "complementary" suggests. (1) Part of the gain is extra capacity. A token-free head closes about 40% of the gap between Base continuation and the full model on ACDM (Table 9). The factorial's query factor mixes observation access with this stage-two head. (2) The paths overlap. The interaction is significantly negative on NURBS and Skeleton, so the paths are sub-additive. (3) Physical consistency gets worse. The per-point correction worsens divergence (0.0465 vs. 0.0283), vorticity, and boundary-normal error (Table 21), which matter to flow practitioners.

To address: (1) add the token-free head as a cell in the factorial; (2) report physical diagnostics on FlowBench; (3) soften "complementary."

Questions for authors
  1. Can the factorial (Table 3) include the parameter-matched token-free head, so that the query factor isolates observation access?

  2. Was writing the correction back into the state tried? If not, why would it be worse?

  3. Is the degradation in Table 21 caused by Δq? For example, how does the corrected output compare with the uncorrected base of the same checkpoint?

  4. Baselines: Did the locally implemented baselines (FieldFormer, S-NOT, FLRONet) get the same tuning budget as GaussAssim? Why is STRIDE not evaluated? Classical methods (Gappy POD, EnKF) are cited but not compared. Even one such baseline on ACDM would show domain readers how much the learned model improves over current practice.

  5. Is the Gaussian state essential, or would a latent token state of equal size behave the same?

  6. I would value one downstream demonstration. Examples: lift and drag estimation from the reconstruction; state estimation for flow control (relevant given the 18.7 ms latency in Table 23); validation on experimental data.

  7. Future work (not required): long-horizon streaming assimilation; spatially coupled or physics-aware regularization of Δq.

9 comments
AuthorsOpenReview ↗

Response to W1: Write-back, longer horizons, and positioning against ANPs

We thank the reviewer for these concrete suggestions. We have added three-seed experiments with trained write-back variants and longer streaming horizons without re-initialization.

1. A write-back variant. We trained two variants that project query corrections from fixed spatial anchors into Gaussian coefficient updates, with feedback strengths 0.1 and 0.3. These updates enter the next state transition. All variants use the same maximum training budget, stopping rules, and validation-based checkpoint selection.

VariantMean error ± SD
Non-writeback0.05005 ± 0.00091
Write-back, strength 0.10.05340 ± 0.00507
Write-back, strength 0.30.05311 ± 0.00297

Training-normalized joint UVP relative error, reported as mean ± sample standard deviation across three training seeds over the evaluation interval specified below.

These results provide direct empirical support for our non-writeback design, which achieves the lowest mean reconstruction error among the evaluated variants under this training and checkpoint-selection protocol.

2. Longer streaming horizons without re-initialization. Following the original ACDM evaluation protocol, we evaluate reconstruction over a 1,097-frame interval (frames 203–1299), scoring every other frame while continuously carrying the recurrent state without re-initialization. Each recording uses a single initialization at its beginning. Evaluation covers all 11 test trajectories and three training seeds.

For non-writeback, the overall and temporal-segment errors are:

Evaluation periodMean error
Overall0.05005
Early0.05845
Middle0.04588
Late0.04583

For each trajectory, these segments are successive portions of one continuous run, not separately initialized evaluations. Mean error decreases from the early to the middle segment and remains stable in the late segment. These results demonstrate that GAUSSASSIM is effective beyond short windows, maintaining stable reconstruction accuracy over the evaluated long streaming horizon.

Current-frame encoding is already competitive on ACDM. original Table 9 (now Table 11) specifically tests query-path history while retaining the M4 state path. The new streaming experiments address the requested extension: sustained recurrence without re-initialization and the effect of feeding query corrections into subsequent states.

3. Clearer positioning against ANPs. ANPs combine a global latent representation with query-conditioned context aggregation. In GAUSSASSIM, the shared representation is an observation-updated Gaussian field state carried into subsequent steps, while query corrections affect only the current output. Our contribution combines this recurrent organization with controlled findings: the four-setting, ten-seed study establishes incremental gains from both paths, and matched readout controls show that letting location determine how observations are combined improves on a shared summary whose decoder already uses coordinates.

Manuscript changes (highlighted in blue). We have sharpened the ANP comparison in Section 2 (pp. 2–3) and added the streaming and write-back results in Section 6.5 (p. 10), Table 7 (p. 10). Appendix G.8 (p. 35), Table 39 (p. 36), and Figure 7 (p. 36) provide the feedback definition, training and checkpoint-selection protocols, per-seed results, and temporal curves over the scoring interval.

AuthorsOpenReview ↗

Response to W2: Extra capacity, sub-additivity, and physical diagnostics

We thank the reviewer for distinguishing these issues. We have extended the factorial comparisons with a token-free head under both assimilation settings and added paired physical diagnostics on FlowBench.

1. Token-free head in the factorial. We compare our query-conditioned correction Q with a capacity-matched token-free head N. N uses query coordinates, geometry, and the Base prediction but does not directly access observation tokens. Within each assimilation setting and seed, N and Q use the same frozen Base.

SettingQ vs. N error reduction: assimilation offQ vs. N error reduction: assimilation on
ACDM4.51%7.47%
FlowBench–NURBS23.70%20.82%
FlowBench–Harmonics24.51%23.80%
FlowBench–Skeleton30.55%26.67%

Relative reduction in three-seed mean reconstruction error, measured against N.

Q outperforms N in all three seeds under both assimilation settings across all four settings. These results directly support the effectiveness of our query-path design: its gains extend beyond those obtained by adding a generic correction head. The revision includes the complete six-cell results, parameter counts, training and checkpoint-selection protocols, and per-seed comparisons.

2. Sub-additivity and “complementary.” We have clarified that “complementary” refers to additional reconstruction gains when the other path is active, as measured in the original ten-seed study, rather than super-additive improvements. The negative interactions on NURBS and Skeleton indicate diminishing combined gains; they do not negate the positive conditional gains. We now use this more precise wording while retaining the distinction between recurrent state assimilation and query-conditioned observation access.

3. Physical diagnostics on FlowBench. We compare predictions before and after query correction using the same full-model checkpoint and test fields. This directly measures the correction’s effect, complementing the comparisons between separately trained models in original Table 21 (now Table 23).

FlowBench familyField reconstruction errorVorticity relative errorNear-body normal-velocity errorDivergence RMSE
NURBS−35.0%−16.1%−2.4%+2.9%
Harmonics−32.0%−14.4%−6.0%+10.2%
Skeleton−34.2%−13.0%−2.5%+4.3%

Changes in three-seed mean errors relative to the same checkpoint’s Base prediction; negative values indicate improvement. FlowBench results in both tables use test frames 228–240.

The correction substantially improves reconstruction and vorticity accuracy, with smaller relative increases in divergence error. Near-body normal-velocity error also decreases on average. These results support the correction’s reconstruction benefits while identifying a specific divergence trade-off, rather than uniform deterioration of physical diagnostics. We explicitly report both the improvements and the divergence increases.

Manuscript changes (highlighted in blue). We have added the token-free factorial extension in Table 4 (p. 8), with complete numerical results and protocols in Appendix G.6 (p. 33), Table 37 (p. 34). Appendix D.4 (p. 26), Table 25 (p. 26), reports FlowBench physical diagnostics. The manuscript now describes conditional reconstruction gains rather than super-additivity.

AuthorsOpenReview ↗

Response to Q1: Parameter-matched token-free head in the factorial

Yes. We have extended the factorial comparison with a capacity-matched token-free head N under both assimilation settings:

State assimilationOriginal no-query configurationToken-free head NQuery-conditioned head Q
OffIncludedAddedIncluded
OnIncludedAddedIncluded

N uses query coordinates, geometry, and the Base prediction without directly accessing observation tokens. Within each assimilation setting and seed, N and Q share the same frozen Base and have closely matched parameter counts. The relevant observation-access contrast is therefore Q versus N, rather than Q versus the original no-query configuration.

As reported in our response to W2, Q reduces mean reconstruction error relative to N by 4.51%–30.55% across ACDM and the three FlowBench families, with improvements in all three seeds under both assimilation settings. This capacity-controlled comparison directly supports the effectiveness of our query-conditioned observation access.

Manuscript changes (highlighted in blue). We have added the separate three-seed, six-cell factorial Table 4 (p. 8) alongside the original ten-seed Table 3. It reports all cell values and capacity-matching conditions. Appendix G.6 (p. 33), Table 37 (p. 34), provides detailed training and checkpoint-selection protocols and per-seed Q/N comparisons.

Reviewer_NE3POpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper tackles the reconstruction of a fluid field (two velocity components and pressure) at any spatial location (x) and a target time (t), given sparse measurements from sensors at known locations collected over (M) snapshots up to and including time (t). The main contribution of the paper is the introduction of a learnable architecture between the input measurements and the field at location (x). They call the architecture GAUSSASSIM, and, to my understanding, it operates as follows: first, a shared encoder maps the information from each sensor to a fixed-size latent vector (token), followed by self-attention layers updating these tokens. Then the tokens are fed into two streams: (1) State Path, where a set of K Gaussian parameters are estimated at the initial time and otherwise refined through a learned propagation network, and the output of the propagation network is passed to an "assimilation" network, which also takes as input the encoded features from the sensor. Then the field is constructed using the Gaussians as basis functions and is also refined spatially by another learned feedforward network. (2) Query Path: This path involves combining the information from past times, queried x, and the output of the State Path to predict a residual correction which, by addition to the State Path's output, defines the final output. They then train this architecture on their chosen benchmark datasets and compare it with other methods!

Critical strengths

I believe the paper deals with a nice problem with possible real-world applications. The experimental validation seems sound, and they perform well on the chosen comparisons.

Critical weaknesses

The presentation of the paper could improve substantially. The current version is hard to follow and sometimes inconsistent. (Please see the questions) The novelty is limited, in my opinion; the components of the model are known, and somehow the novelty lies in their combination. The paper uses complex language to describe simple ideas, e.g., phrases like "near-body measurements" in the abstract could be replaced with simpler descriptions; this pattern happens throughout the paper.

Questions for authors
  1. In equation 1, what is happening to s? I guess you should iterate over s also to obtain the full measurements?
  2. In Figure 1, where is the input encoder? Also, it is written there that O contains snapshots of q, while in part 4, it is written that it contains the locations of the sensors as well.
  3. Do you have any interpretability results for the whole final mapping? Specifically, could you visualise or quantify how each sensor at each time is affecting the final output?
  4. For the presentation, I strongly suggest the improvement of Figure 1. Then, you could have a paragraph (or a few lines) about each module in that figure in the methods section, with equations describing the parameters and the input and output of each module.
6 comments
AuthorsOpenReview ↗

Response to W1: Contribution and presentation

We thank the reviewer for recognizing the relevance of the problem and the strength of the experimental evaluation.

Contribution. Our matched comparison establishes a specific design finding for recurrent flow reconstruction. Both readouts use the query location to predict the field. The shared-summary readout combines observations in the same way for every location; our query-conditioned readout lets each location determine how those observations are combined.

With the same frozen Base, the same encoder and residual-decoder architectures, closely matched parameter counts, and matched training protocols, query-conditioned aggregation reduces error by 6.36% on NURBS and 7.48% on ACDM Inc. Thus, even after observations have been assimilated into the state, location-specific observation aggregation improves on a shared summary whose decoder already uses coordinates.

This finding complements our architectural contribution: GAUSSASSIM separates observation updates carried forward in a recurrent Gaussian state from observation-based corrections confined to the current output. Our four-setting, ten-seed study shows that each path further reduces mean reconstruction error when the other is active. Together, these results establish the value of both the dual-path organization and the query-dependent reading of observations within this framework.

Presentation. We have made the computation and contributions easier to follow through targeted revisions:

  • Figure 1: now explicitly shows the shared sensor encoder, its inputs, and its connections to both paths.
  • Eq. 1: now specifies both time and sensor indices and uses consistent observation definitions throughout.
  • Method: now follows the figure’s computational order, defining module inputs, outputs, dimensions, and shared features alongside the equations.
  • Language: now uses direct descriptions, such as “measurements from sensors near the object.”

Manuscript changes (highlighted in blue). Figure 1 (p. 3) and Sections 3–4 (pp. 3–5) incorporate these clarifications, with terminology simplified throughout the manuscript.

AuthorsOpenReview ↗

Response to Q1: Sensor indexing in Eq. 1

We thank the reviewer for pointing out this ambiguity. In the first paragraph of Section 3, immediately before Eq. 1, we introduce SS sensors with coordinates Xs=xss=1SX_s=\\{x_s\\}_{s=1}^{S}. The observation window includes measurements from all SS sensors at each of the MM sampled times. We have made both index ranges explicit in Eq. 1:

Y_{t-M+1:t}=\left\\{y_{\tau,s}=q_{\tau}(x_s)\right\\}_{\tau=t-M+1,\ldots,t;\\;s=1,\ldots,S}.

Each yτ,s∈R3y_{\tau,s}\in\mathbb{R}^{3} contains the two velocity components and pressure measured at sensor ss and time τ\tau. The complete window therefore contains M×SM\times S measurement vectors.

Manuscript changes (highlighted in blue). We have updated Eq. 1 and its accompanying explanation in Section 3 (p. 3) to explicitly show both indices.

AuthorsOpenReview ↗

Response to Q2: Input encoder and observation definition

We thank the reviewer for pointing out these ambiguities.

The input encoder is the shared sensor encoder defined in Section 4.1. We have made it explicit in the revised Figure 1, between the measurements and the two paths. It processes each snapshot separately, using sensor coordinates, UVP measurements, and available local geometry to produce per-sensor features. The state path uses these features sequentially; the query path further incorporates temporal information to construct its observation bank.

Each snapshot OτO_\tau contains sensor locations paired with their measurements: (xs,yτ,s)(x_s,y_{\tau,s}), where yτ,s=qτ(xs)y_{\tau,s}=q_\tau(x_s) and s=1,…,Ss=1,\ldots,S. The previous caption’s phrase “snapshots of q” abbreviated this definition and omitted the locations. The revised input label explicitly states “Sensor locations + UVP measurements.” The illustrated background fields provide visual context only; they are not model inputs.

Manuscript changes (highlighted in blue). We have replaced Figure 1 (p. 3) and updated its caption to match the observation and encoder definitions in Section 4.1 (p. 4).

Reviewer_KhF6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how sparse observations should be integrated into recurrent models for continuous spatiotemporal field reconstruction. The authors propose GAUSSASSIM, which represents the evolving field using a recurrent state composed of anisotropic Gaussian primitives. The model incorporates observations through two complementary paths: a state path that assimilates observations into the recurrent Gaussian representation, and a query-conditioned path that directly attends to time-indexed observation tokens to predict coordinate-specific corrections without modifying the recurrent state. The two paths are trained in stages, with the base model optimized first and then frozen while the query-conditioned correction module is trained. Experiments on ACDM Inc and FlowBench evaluate the approach against several baselines and include ablations of the two paths, temporal observation inputs, and alternative readout mechanisms. The results suggest that direct query-conditioned access to observations can complement recurrent state assimilation and improve reconstruction accuracy in the evaluated settings.

Critical strengths

Clear problem formulation and well-motivated design. The paper identifies a meaningful limitation of recurrent state-based reconstruction: observations compressed into a latent state may not remain directly accessible to individual spatial queries. The distinction between state assimilation and query-conditioned observation access is intuitive and provides a clear motivation for the proposed architecture.

A complementary dual-path architecture. GAUSSASSIM combines recurrent Gaussian-state updates with direct cross-attention to time-indexed observation tokens. The state path supports temporal information accumulation and state propagation, while the query path allows different spatial coordinates to retrieve different observational evidence. This separation offers a principled way to combine persistent latent representations with direct access to observations.

Systematic ablation studies. The paper evaluates the contributions of the state and query paths through factorial ablations and investigates alternative temporal inputs and observation aggregation mechanisms. These experiments go beyond reporting aggregate performance and provide evidence for the proposed design choices.

Critical weaknesses

(1) Lack of diffusion-based baselines. The paper lacks comparisons with relevant diffusion-based PDE methods, such as Guided Diffusion Sampling on Function Spaces with Applications to PDEs (Yao et al., 2026) and DiffusionPDE: Generative PDE-Solving under Partial Observation (Huang et al., 2024). Including suitable comparisons would better establish the advantages of GAUSSASSIM. (2) Limited generalization to sensor configurations. The model's performance may depend on the specific observation locations used during training. It is unclear whether it can generalize to randomly sampled sensor locations or different predefined sensor configurations. Without such generalization, changing sensor placements may require retraining, limiting the method's practical applicability.

Questions for authors

(1) How does your method address the cold-start problem at the initial timesteps, when only sparse near-body observations are available and insufficient temporal context exists to accurately reconstruct the global flow field? (2) How well does the query-based observation pathway perform at super-resolution, given its ability to query the reconstructed field at arbitrary spatial locations? Can this zero-shot generalization capability benefit from the initial Gaussian representation.

4 comments
AuthorsOpenReview ↗

Response to W1: Diffusion-based PDE baselines

We thank the reviewer for this helpful suggestion. In Section 2 of the revised manuscript, we have added a discussion of diffusion-based PDE reconstruction, including DiffusionPDE and FunDPS. These methods use learned distributions of fields and iterative guided sampling to recover solutions from partial observations.

Task setting. GAUSSASSIM addresses supervised flow-field reconstruction under continuing sensing. Its inputs are sparse sensor measurements from a short time window that includes the target time, and its output is the dense velocity–pressure field at that time. Our study examines how these measurements should contribute to both the evolving state estimate and the reconstruction at each spatial location. Diffusion-based methods offer a complementary approach with different training objectives and inference procedures, requiring adaptation to our observation window and reconstruction task.

Additional ACDM comparison. To further evaluate GAUSSASSIM against these methods, we adapted the official DiffusionPDE and FunDPS implementations to ACDM Inc with S16/M4 observations. All three methods receive the same measurements at 16 near-cylinder sensors over four consecutive frames and use the same trajectory splits and training normalization. Both diffusion priors were trained on all 70 training trajectories. Evaluation uses a common subset of 220 test fields, with 20 fields selected by a fixed temporal rule from each of the 11 test trajectories.

MethodFull field ↓16 observed points ↓Near cylinder ↓Far downstream ↓
GAUSSASSIM0.018560.028140.014820.02733
DiffusionPDE (ACDM adaptation)0.079190.037260.016950.13872
FunDPS (ACDM adaptation)0.119420.001320.065500.17902

Test results from one training seed (123), averaged over 220 fields across all 11 test trajectories. Full-field errors use the training-normalized joint velocity–pressure relative L2L^2 metric in Eq. (3) and Table 2. Regional errors apply the same formula within each indicated region. Far downstream denotes the final quarter of the streamwise grid.

Results and spatial diagnostics. GAUSSASSIM achieves the lowest full-field test error. FunDPS fits the observed points most closely, with an error of 0.00132, but its far-downstream error reaches 0.17902. DiffusionPDE approaches GAUSSASSIM near the cylinder (0.01695 versus 0.01482), while the difference is larger downstream (0.13872 versus 0.02733). Thus, closer agreement at the sensors does not necessarily yield more accurate recovery of the unobserved field. GAUSSASSIM achieves lower errors over the full field and in both spatial regions under the evaluated setting.

We have added these supplementary results and spatial diagnostics to Appendix G.5 (p. 33), Table 36 (p. 34), which also documents the method adaptations, evaluation settings, computational budgets, and limitations. Section 5.2 (p. 6) now summarizes the comparison and directs readers to the appendix for the complete settings and results.

AuthorsOpenReview ↗

Response to W2: Generalization across sensor configurations

We thank the reviewer for raising this practical concern.

Original sensing setting. Our primary experiments consider sensors fixed relative to the body, motivated by measurements at prescribed locations such as airfoil surface-pressure sensing. This fixed-location assumption is also used in Willcox (2006), already cited in our introduction, which studies unsteady flow reconstruction from limited surface-pressure measurements.

In ACDM Inc, the same 16 near-cylinder locations are used across Reynolds trajectories and data splits. In FlowBench, each geometry has its own sensor layout, held fixed across Reynolds numbers, time, and data splits (Appendices A.2–A.3). We also recognize the practical value of using a trained model across alternative sensor configurations.

Additional random-layout experiment. We therefore conducted a three-seed experiment on ACDM Inc with random sensor layouts during training. For each training window, 16 distinct locations are sampled from the first fluid-cell layer adjacent to the cylinder, with the same locations used across its four frames. Both training stages use varying layouts while retaining the original architecture and 30+15-epoch schedule. The fixed-layout models use the same total training epochs.

At test time, each seed uses one frozen checkpoint across all six configurations, without configuration-specific retraining or fine-tuning. The configurations comprise the original layout, two boundary-shifted layouts, and three random layouts. Each is evaluated on all 6,039 test fields.

Test configurationFixed-layout training ↓Random-layout training ↓
Original layout0.01954 ± 0.002610.03503 ± 0.00165
Boundary shift −10.56875 ± 0.122380.03673 ± 0.00128
Boundary shift +10.37935 ± 0.040450.03355 ± 0.00130
Random layout 10.52335 ± 0.010020.03372 ± 0.00119
Random layout 20.51915 ± 0.157010.04048 ± 0.00477
Random layout 31.00131 ± 0.460280.04905 ± 0.00648

Training-normalized joint velocity–pressure relative L2L^2 error, using the metric from Table 2. Mean ± sample SD across seeds 123/456/789, with 6,039 test fields per configuration per seed. Boundary shifts move each sensor by one position along the ordered near-body candidate set.

Results and deployment implications. Random-layout training reduces mean error by 91.2%–95.1% across the five alternative configurations, with improvements in all three seeds for every configuration. The resulting errors range from 0.03355 to 0.04905. These results show that a model trained with varying layouts can accommodate the tested near-body configurations without retraining for each placement.

Random-layout training also increases original-layout error from 0.01954 to 0.03503. This quantifies the trade-off between accuracy at the original sensor locations and generalization across configurations. The original architecture supports both training settings, allowing the choice to reflect whether deployment uses a stable installation or multiple sensor arrangements.

We have clarified the original sensing assumptions in Section 5.1 (pp. 5–6) and added this cross-configuration comparison to Appendix G.3 (p. 31), Table 33 (p. 32), including the layout construction, training settings, and complete results.

AuthorsOpenReview ↗

Response to Q1: Cold-start reconstruction

We thank the reviewer for raising this practical question.

Startup procedure. GAUSSASSIM initializes its Gaussian state from the first sparse snapshot through a learned initialization network. The subsequent three snapshots refine this state through propagation and assimilation before the standard four-snapshot reconstruction. This requires neither a dense initial flow field nor a state carried over from previous windows.

Task setting and evaluation. Our primary protocol uses complete four-snapshot windows and evaluates post-startup target fields (Sections 3-5). To examine incomplete-history operation, we evaluated the same pretrained models with one to four available snapshots on the same 220 ACDM target fields across 11 test trajectories, using three training seeds.

Available snapshotsFull-field error ↓\downarrow
10.61838±0.149460.61838 \pm 0.14946
20.46538±0.148890.46538 \pm 0.14889
30.23852±0.077140.23852 \pm 0.07714
40.02019±0.002580.02019 \pm 0.00258

Mean ±\pm sample standard deviation across three seeds, using the training-normalized joint velocity-pressure relative L2L^2 metric in Eq. (3). Targets are identical across history lengths; these are not successive startup timesteps.

Results and scope. Error decreases as more history becomes available, reaching 0.02019 with the complete window. These results quantify the behavior of the four-snapshot-trained model under incomplete history; they do not establish equally accurate reconstruction from the first observation.

We have clarified the complete-window scope of our reported accuracy and added these diagnostics to Appendix G.4 (p. 32). Improving early predictions through variable-length-history training is a direction for future work.

AuthorsOpenReview ↗

Response to Q2: Spatial querying and zero-shot resolution transfer

Task scope. Our work does not specifically target super-resolution. GAUSSASSIM addresses velocity–pressure field reconstruction from sparse sensor measurements.

Mechanism. The Gaussian state provides a continuous spatial representation, and the query-conditioned readout evaluates the reconstructed field at arbitrary coordinates within the fluid domain. Consequently, the output representation is not tied to a fixed grid resolution: the same trained model can be queried on a finer grid while keeping its parameters and input observations unchanged.

Evaluation. We evaluated models trained at 512×128 directly against native 1024×256 ground truth, without retraining. The comparison uses the same ten geometry–Reynolds cases per FlowBench family, one fixed test time per case, and three training seeds.

FamilyError at 512×128 ↓Error at 1024×256 ↓
NURBS0.07273 ± 0.004070.07254 ± 0.00406
Harmonics0.08864 ± 0.009150.08846 ± 0.00914
Skeleton0.09631 ± 0.007460.09588 ± 0.00744

Mean ± sample standard deviation across three seeds, using training-normalized joint velocity–pressure relative L2L^2 error on the valid evaluation support at each resolution. Fine-grid evaluation additionally excludes points whose coarse-grid sampling neighborhood crosses the solid boundary.

The reconstruction errors remain numerically close when the output resolution is doubled along each spatial dimension. These results support zero-shot reconstruction on a finer output grid without a substantial change in error in the tested setting, consistent with the continuous spatial formulation of our method.

We have included this evaluation and its protocol details in Appendix G.4 (p. 32), Table 34 (p. 32).

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

This paper proposes GAUSSASSIM, continuous spatial field reconstruction method from discrete-time sparse sensor measurements. Estimation of the target-time field at a query coordinate is done by passing observation into two separate path; state path, which corresponds to temporal evolution of Gaussian recurrent latent state and query path, which corresponds to generating residual correction at a query point w.r.t attended field decoded from the recurrent state. The paper conducts experiments on ACDM Inc and FlowBench to evaluate: 1. reconstruction accuracy compared to existing methods, 2. generalization to unseen Reynolds numbers, 3. ablation on the roles of the state and query paths, 4. the contribution of past observations, and 5. the benefit of query-conditioned observation aggregation compared to a shared observation representation.

Critical strengths

The paper proposed flow field reconstruction method that separates the state path for assimilation and query correction, allowing two path to complement each other. Overall design decisions are reasonable as state path maintains recurrent information while query path directly evaluate on corrdinate-specific observation. The paper provided extensive experiments which is enough to validate the reciprocal role of each components and effectiveness compared to baseline method.

Critical weaknesses

Although motivation and method were clearly stated, I found it difficult to follow the experimental sections, as many details needed to understand the experiments were provided in the appendix. Also, some background knowledge of flow field reconstruction seemed necessary to fully understand the experimental settings and the purpose of each experiment. Some of the notation isn't defined explicitly, which hurts readability - see my questions below. The paper emphasizes state and query path separation is important, particularly by not feeding the correction back into the recurrent state. While the design decision seems reasonable but I am not fully convinced why this is necessary. Providing empirical evidence to support this design choice would help strengthen the argument

Questions for authors
  1. Could you define cRec_{Re}, mcm_c, scs_c, PmP_m, PtP_t, and PqP_q in the paper? What are the output dimensions of the networks PtP_t, PmP_m, and PqP_q? Do they share any features?
  2. Why do we need hth_t if ata_t already provides the UVP field? Can’t the decoder generate different outputs at different query coordinates even without hth_t? I’m wondering how the roles of ata_t and hth_t differ and what kind of information each represents.
  3. Can the model generate a field at an arbitrary time between observed time steps, rather than only at the observed times?
  4. Could you justify why the query correction is not used for state assimilation? What happens if feedback from the correction is allowed? Is there empirical evidence that separating query correction from state evolution is preferable to incorporating the corrected information into subsequent state updates?
  5. For the first experiment, why did you choose frames [60, 240) for the periodicity screening? How does the model perform on frames before and after this interval, where periodicity may be weaker?
6 comments
AuthorsOpenReview ↗

Response to W1: Experimental clarity and non-writeback design

We thank the reviewer for these suggestions.

Experimental clarity. We have made the experimental sections self-contained by stating each experiment's purpose, inputs, comparisons, and evaluation setting before presenting its results. Notation, network dimensions, and shared features are explicitly defined in the main text.

Non-writeback design and evidence. Observations already update the recurrent state through assimilation; the query path additionally corrects the current output. Feeding this correction back requires mapping it into Gaussian state updates.

We tested this alternative using three training seeds in a streaming setting with 1,300 successive observations and one initialization per trajectory. Once four observations are available, corrections at 256 fixed spatial anchors are approximately projected onto Gaussian coefficient updates using ridge regression. These updates are scaled by 0.1 or 0.3 and incorporated before the next recurrent step. Each variant is trained with its corresponding feedback rule.

All variants follow the same validation-based stopping rule, with a maximum of 240 epochs. Selecting the checkpoint with the lowest validation error from each completed run gives:

VariantTest error
No feedback0.05005 ± 0.00091
Feedback strength 0.10.05340 ± 0.00507
Feedback strength 0.30.05311 ± 0.00297

Training-normalized joint UVP relative L2L^2 error, mean ± sample standard deviation across three seeds. Each model is evaluated on 6,039 fields from 11 test trajectories, covering frames 203, 205, ..., 1299. This streaming comparison is separate from the main four-frame-window experiments.

Under this stopping and checkpoint-selection protocol, non-writeback achieves the lowest mean test error. The results support its use for streaming reconstruction without an additional field-to-state projection.

Manuscript changes (highlighted in blue). Sections 3–4 (pp. 3–5) clarify notation, network dimensions, and shared features. Section 5 (p. 5) introduces experimental settings and purposes. Section 6.5 (p. 10), Table 7 (p. 10) presents the streaming feedback comparison; Appendix G.8 (p. 35), Table 39 (p. 36), and Figure 7 (p. 36) provide the detailed protocols and per-seed results.

AuthorsOpenReview ↗

Response to Q1: Notation, dimensions, and feature sharing

We thank the reviewer for requesting these clarifications. The Reynolds condition and shared sensor features are introduced in Sections 3 and 4.1, while the architecture widths are reported in Appendix B.1. We agree that these definitions and dimensional relationships should be presented together more explicitly.

Normalization and conditioning. For each channel c∈u,v,pc\in\\{u,v,p\\}, mcm_c and scs_c denote the training-data mean and standard deviation: q~c=qc−mcsc.\widetilde q_c=\frac{q_c-m_c}{s_c}. These statistics remain fixed during validation and testing. The scalar cRe=log⁡Re−log⁡100log⁡1000−log⁡100c_{\mathrm{Re}}=\frac{\log Re-\log100}{\log1000-\log100} is used only for geometry-OOD conditioning, not in the primary experiments.

Network dimensions. The projections in the primary models are:

NetworkInput and operationOutput dimension
PmP_mLinear projection of sensor feature eτse_{\tau s}96
PtP_tMLP embedding of scalar relative time offset δτ\delta_\tau96
PqP_qMLP embedding of Fourier query coordinates and available geometry/conditioning features96

The shared sensor encoder produces 96 features per sensor. The observation-token encoder preserves the MS×96MS\times96 shape, giving 64×9664\times96 for M=4,S=16M=4,S=16, excluding the batch dimension. The correction head CψC_\psi combines the retrieved 96-dimensional feature with the three-channel Base prediction and outputs a three-channel UVP correction (99→399\rightarrow3).

Feature sharing. Both paths share the sensor encoder and its output features. The query correction additionally receives the state-decoded Base prediction. Their downstream state-update and query-attention modules have separate parameters. Likewise, Pm,Pt,PqP_m,P_t,P_q have independent weights despite their common output width. During query-stage training, the shared encoder remains frozen with the Base model.

Manuscript changes (highlighted in blue). Section 3 (p. 3) explicitly identifies mc,scm_c,s_c beside normalization. Section 4.1 (p. 4) clarifies feature and parameter sharing. Section 4.3 (p. 5) specifies projection and correction-head dimensions symbolically; Section 5.1 gives their experimental values, with full configuration details in Appendix B.1.

AuthorsOpenReview ↗

Response to Q2: Why use h_t in addition to a_t?

Yes, the decoder can generate different outputs at different query coordinates even without hth_t. The Gaussian weights vary with the query coordinate, and the decoder also receives coordinate features. We use hth_t to improve reconstruction, not to enable spatial variation.

What do ata_t and hth_t represent? For each Gaussian, ata_t contains three UVP coefficients. Combining these coefficients with the position-dependent Gaussian weights produces a field estimate. By contrast, hth_t is a 96-dimensional learned vector per Gaussian. It provides additional information for updating the state and refining the decoded field, rather than directly specifying another UVP field. Both ata_t and hth_t are updated during recurrence; individual entries of hth_t are learned through reconstruction training rather than assigned specific physical quantities.

Does using hth_t improve reconstruction? We tested this on ACDM using three training seeds. For each seed, we kept the recurrent model fixed and fine-tuned copies of its decoder for 15 epochs, either using both ata_t and hth_t or disabling direct access to one of them at readout. Coordinate inputs remained available, and query correction was disabled in all variants.

Decoder inputsReconstruction error
Both ata_t and hth_t0.02074 ± 0.00205
Without direct use of ata_t0.02611 ± 0.00201
Without direct use of hth_t0.03956 ± 0.00064

Normalized joint UVP relative L2L^2 error, mean ± sample standard deviation across three seeds, evaluated on the same 6,039 test fields from 11 trajectories. The frozen recurrent state retains both ata_t and hth_t; only their direct use by the decoder changes.

Using both gives lower error in every seed. Thus, the decoder can produce spatially varying predictions without hth_t, but using hth_t makes those predictions more accurate in this comparison.

Manuscript changes (highlighted in blue). Section 4.2 (p. 4) clarifies what ata_t and hth_t represent and how each is used. Appendix G.1 (p. 31), Table 31 (p. 31), reports the decoder comparison and experimental settings.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary for the Area Chair and Reviewers

Overall Response

We thank the Area Chair and all four reviewers for their careful assessment and constructive feedback. We have revised the manuscript and conducted additional experiments to clarify the contribution, test the key design choices, and improve the presentation.

1. Clarified contribution and design finding. GAUSSASSIM separates observation-driven updates carried forward in a recurrent Gaussian state from observation-based corrections to the current output. Our contribution includes both this architecture and controlled findings about its design. The four-setting, ten-seed study establishes additional reconstruction gains from each path when the other is active. The matched readout comparison identifies a more specific finding: both readouts use the prediction location, but accuracy improves when that location also determines how observations are aggregated. This reduces error by 6.36% on NURBS and 7.48% on ACDM Inc relative to the tested shared-summary readout.

2. Further validation of the method. We added capacity-matched token-free controls under both assimilation settings, trained write-back alternatives, and streaming experiments without window-wise re-initialization. These experiments strengthen the evidence for direct observation access and support our non-writeback design over the evaluated streaming horizon. An equal-sized latent-token comparison additionally evaluates the Gaussian state representation.

3. Broader empirical evaluation. We added ACDM comparisons with Gappy POD, DiffusionPDE, and FunDPS, together with evaluations of sensor-layout transfer, incomplete histories, finer spatial queries, and same-checkpoint physical diagnostics. Sensor-by-time interventions further quantify how individual observations affect the complete model’s final prediction.

4. Improved clarity and organization. We revised Figure 1, made observation indices and module dimensions explicit, and clarified each experiment’s purpose, comparison, and evaluation scope. The main text presents the central findings, while the appendix provides the supporting protocols and detailed results. Revisions are highlighted in blue, and the point-by-point responses identify their locations.

Together, these revisions make the architectural contribution, the resulting design knowledge, and their empirical support more explicit. We appreciate the reviewers’ suggestions and the Area Chair’s consideration of the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.