acceptodds
Under review as a conference paper at ICLR 2027

Shaping the Wind: Nested Potentials for Kinematically Admissible Urban Wind Prediction

18 views

Abstract

Predicting transient urban winds is fundamental to understanding urban microclimates and designing climate-resilient cities. Building-resolving large-eddy simulation produces detailed incompressible urban wind fields at substantial computational cost for each layout. Neural surrogates offer a faster alternative by learning to predict the evolution of velocity fields. However, minimizing velocity prediction error does not guarantee local mass conservation and wall impermeability, which together define kinematic admissibility. This limitation stems from an unconstrained output representation: geometry conditioning guides predictions but does not restrict them to admissible velocity fields. Correcting boundary violations in these outputs changes the flux balance in adjacent fluid cells and may consequently compromise local mass conservation. To address the challenge, we propose Sculpt, a nested potential framework that builds the coupled, geometry-dependent constraints directly into its parameterization. This nested parameterization generates divergence-free velocity updates through the discrete curl of a volume vector potential on the native three-dimensional staggered grid. A shared scalar potential constrains the vector potential's boundary values so that the same operator also enforces impermeability, without a per-step pressure projection. Because backpropagation through this curl attenuates large-scale gradient signals, we parameterize the volume potential at multiple resolutions to better capture large-scale flow structures. We introduce UrbanWindFlow, an LES dataset spanning urban morphologies and inflow conditions, to evaluate accuracy and kinematic admissibility together. Across all three evaluation settings, Sculpt consistently outperforms both the reduced-order and the neural baselines, reducing mean squared error by – while maintaining constraint residuals near numerical precision.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Wind Field Prediction, Neural Surrogate, Physical Constraints

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_TScWOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Sculpt, an autoregressive urban-wind surrogate whose output parameterization enforces discrete incompressibility and wall impermeability on a staggered grid. A volume vector potential yields divergence-free velocity increments, while a shared scalar potential and harmonic modes constrain boundary values so the same curl readout has zero wall-normal velocity. A multiresolution potential parameterization addresses the scale-dependent gradients induced by the curl.

Critical strengths
  • The evaluation is good. It covers reduced-order, convolutional, spectral, and attention baselines, as well as targeted constraint ablations and a post-projection comparator
  • The core construction seems to be well motivated. The authors connect the discrete operators directly to the two advertised guarantees rather than relying on soft penalties.
Critical weaknesses
  • I am not sure if the statistical analysis is sufficient here in this paper. Reported deviations are over overlapping test rollouts, not independent training runs. It would be better to provide repeated training seeds and paired sequence-level results, especially for the smaller ablation gains.
  • “Geometry generalization” is a bit overstated. Held-out layouts and morphology classes test transfer within one procedural, grid-aligned generator.
  • Avoiding pressure projection is presented as an efficiency advantage, yet no matched latency, throughput, memory, or parameter comparison is reported. The rollout guarantee also requires an admissible initial field; can the authors clarify how imperfect initial fields would be handled.
Questions for authors
  • How far do the held-out layout and morphology results support transfer beyond this grid-aligned procedural generator and fixed periodic domain?
  • Is there evidence on measured city layouts or nonperiodic horizontal boundaries? For curved boundaries, cut cells, and other discretizations, what changes would be needed to retain the guarantee?
  • How stable are the gains over U-Net and P3D across independent training seeds and held-out sequences, using paired sequence-level comparisons?
  • What are matched per-step latency, rollout time, peak memory, parameter count, and training cost relative to U-Net, P3D, and the post-projection baseline?
  • Beyond the shared inputs, splits, epochs, and checkpoint rule, how were pushforward training, hyperparameter tuning, and model capacity handled for each baseline? Were capacity or tuning-budget sensitivity checks performed?
7 comments
AuthorsOpenReview ↗

Author Response to Reviewer TScW (W1, W2 & Q1)

We thank the reviewer for the comments. Please find our responses to your concerns below.

W1: Statistical Analysis

Response to W1:

The mean and standard deviation reported in the manuscript summarize predictions across different test sequences and frames. As requested, we report the mean and sample standard deviation across three training seeds for the main experiments under the unseen-layout and unseen-morphology settings, and for the ablation study under the unseen-layout setting. All MSE values are in units of 10−310^{-3}. We will report the complete experimental results in the revised manuscript.

Main experiments — Unseen layouts

\hline \text{Method} & \text{MSE@1 }\downarrow & \text{MSE@4 }\downarrow & \text{MSE@8 }\downarrow & \text{MSE@16 }\downarrow \\\\ \hline \text{Dil-ResNet} & 1.14 \pm 0.38 & 10.52 \pm 2.73 & 36.53 \pm 6.60 & 105.82 \pm 32.31 \\\\ \text{TFNO} & 14.82 \pm 6.01 & 30.70 \pm 6.10 & 56.85 \pm 12.61 & 109.66 \pm 32.28 \\\\ \text{AFNO} & 10.49 \pm 1.76 & 29.01 \pm 1.24 & 44.00 \pm 1.85 & 63.13 \pm 2.05 \\\\ \text{AViT} & 7.64 \pm 0.25 & 17.88 \pm 0.26 & 33.53 \pm 0.92 & 59.31 \pm 1.32 \\\\ \text{P3D} & 1.08 \pm 0.06 & 6.67 \pm 0.72 & 19.11 \pm 2.10 & 44.52 \pm 1.29 \\\\ \text{U-Net} & 0.54 \pm 0.07 & 3.00 \pm 0.26 & 8.47 \pm 0.51 & 24.40 \pm 1.33 \\\\ \text{PCNO} & 21.11 \pm 0.01 & 34.69 \pm 0.08 & 40.97 \pm 0.25 & 49.48 \pm 0.43 \\\\ \text{clawFNO} & 17.28 \pm 0.18 & 27.33 \pm 0.26 & 33.74 \pm 0.64 & 44.68 \pm 2.01 \\\\ \text{Sculpt} & 0.21 \pm 0.01 & 0.91 \pm 0.15 & 3.35 \pm 0.48 & 14.59 \pm 1.14 \\\\ \hline \end{array}$$ **Main experiments — Unseen morphologies** $$\begin{array}{l|cccc} \hline \text{Method} & \text{MSE@1 }\downarrow & \text{MSE@4 }\downarrow & \text{MSE@8 }\downarrow & \text{MSE@16 }\downarrow \\\\ \hline \text{Dil-ResNet} & 1.76 \pm 0.62 & 15.15 \pm 3.87 & 48.43 \pm 7.86 & 128.20 \pm 36.53 \\\\ \text{TFNO} & 19.10 \pm 7.71 & 41.24 \pm 7.48 & 73.37 \pm 16.26 & 125.24 \pm 33.84 \\\\ \text{AFNO} & 13.96 \pm 2.17 & 38.64 \pm 1.38 & 59.62 \pm 2.04 & 91.07 \pm 4.56 \\\\ \text{AViT} & 10.57 \pm 0.21 & 27.82 \pm 0.80 & 53.25 \pm 1.95 & 88.56 \pm 2.80 \\\\ \text{P3D} & 1.67 \pm 0.09 & 10.10 \pm 0.75 & 28.44 \pm 1.76 & 61.66 \pm 1.44 \\\\ \text{U-Net} & 0.82 \pm 0.10 & 4.91 \pm 0.57 & 13.86 \pm 1.44 & 37.28 \pm 3.52 \\\\ \text{PCNO} & 25.80 \pm 0.01 & 44.61 \pm 0.25 & 54.52 \pm 0.60 & 68.09 \pm 1.08 \\\\ \text{clawFNO} & 21.40 \pm 0.16 & 36.39 \pm 0.14 & 47.06 \pm 0.70 & 64.31 \pm 3.10 \\\\ \text{Sculpt} & 0.30 \pm 0.02 & 1.62 \pm 0.23 & 6.00 \pm 0.60 & 24.34 \pm 0.78 \\\\ \hline \end{array}$$ **Ablation study — Unseen layouts** $$\begin{array}{l|cc} \hline \text{Variant} & \text{MSE@1 }\downarrow & \text{MSE@16 }\downarrow \\\\ \hline \text{Sculpt} & 0.21 \pm 0.01 & 14.59 \pm 1.14 \\\\ \text{w/o null-space }\Psi_g & 0.41 \pm 0.02 & 31.17 \pm 0.22 \\\\ \text{w/ soft penalty} & 17.39 \pm 7.12 & 115.26 \pm 13.76 \\\\ \text{w/ post-projection} & 0.36 \pm 0.02 & 25.26 \pm 0.82 \\\\ \text{w/o boundary }G & 0.27 \pm 0.02 & 19.32 \pm 2.50 \\\\ \text{w/o increment }\Delta\mathbf{u} & 11.31 \pm 6.74 & 3349.35 \pm 914.49 \\\\ \text{w/o multi-scale }\mathbf{A} & 0.21 \pm 0.01 & 15.35 \pm 0.75 \\\\ \text{w/o learned fusion} & 0.24 \pm 0.02 & 16.65 \pm 1.67 \\\\ \hline \end{array}$$ These results show that our model’s performance and the gains demonstrated by the ablations are consistent across seeds. ------ **W2 & Q1: Scope of Geometry Generalization** **Response to W2 & Q1:** In urban wind modeling, previous studies [1,2,3] have similarly evaluated generalization using unseen geometries generated within a shared procedural framework. For example, [4] evaluated unseen urban geometries from the same parametric generator, while [5] tested unseen synthetic urban layouts generated under common geometric rules. Consistent with these established evaluation practices, our experiments assess generalization to both unseen layouts and held-out morphology classes. Therefore, we believe geometry generalization appropriately describes the evaluated capability, without implying unrestricted transfer across arbitrary geometric representations or domains.
AuthorsOpenReview ↗

Author Response to Reviewer TScW (W2 & Q1, W3)

Response to W2 & Q1 (continued):

To Q1: Our experiments provide evidence of geometric transfer at two levels: held-out layouts test generalization to unseen building arrangements, while held-out morphology classes further evaluate transfer to geometric patterns absent from training. These results demonstrate that Sculpt can generalize beyond the specific geometries encountered during training, rather than merely memorizing individual layouts. However, both evaluations remain within the same procedural generation framework, grid-aligned representation, and periodic domain. Transfer to held-out layouts or nonperiodic domains has not yet been evaluated and remains an important direction for future work, as discussed in the Limitations and Future Work section.

References

[1] GeoWind2Plan: Mission-Time 3D Urban Wind Prediction for Energy-Efficient UAV Planning. NeurIPS, 2026.

[2] Generalization of urban wind environment using Fourier neural operator across different wind directions and cities. arXiv preprint arXiv:2501.05499 (2025).

[3] Building a general and data-efficient convolutional neural network-based model for fast urban flow estimation. Building and Environment (2026): 114488.

[4] A convolutional neural network for steady-state flow approximation trained on a small sample size. Atmosphere 14.9 (2023): 1462.

[5] From large-eddy simulations to deep learning: A U-net model for fast urban canopy flow predictions. Sustainable Cities and Society (2025): 107005.


W3: Computational Efficiency and Imperfect Initial Fields

Response to W3:

Thank you for raising these two points. We have incorporated a computational complexity analysis into the revised manuscript to provide a more direct efficiency comparison. Specifically, we use the existing "w/ post-projection" ablation as the reference and evaluate two pressure-Poisson solvers: the original Conjugate Gradient (CG) method [1] and a GPU-based Fast Fourier Transform–preconditioned Conjugate Gradient (FFT-PCG) implementation [2]. Both projection variants share the same trained velocity predictor and introduce no additional trainable parameters. The reported inference costs account for the complete prediction process, including velocity prediction and pressure projection.

\hline \text{Method} & \begin{array}{c}\text{Unseen layouts: }\\\\\text{Train (s/epoch)}\end{array} & \begin{array}{c}\text{Unseen layouts: }\\\\\text{Inference (s/step)}\end{array} & \begin{array}{c}\text{Unseen morphologies: }\\\\\text{Train (s/epoch)}\end{array} & \begin{array}{c}\text{Unseen morphologies: }\\\\\text{Inference (s/step)}\end{array} & \begin{array}{c}\text{Unseen inflows: }\\\\\text{Train (s/epoch)}\end{array} & \begin{array}{c}\text{Unseen inflows: }\\\\\text{Inference (s/step)}\end{array} \\\\ \hline \text{PALM (LES; 16 MPI processes)} & \text{—} & 4.32 & \text{—} & 4.32 & \text{—} & 4.32 \\\\ \text{Dil-ResNet} & 457.93 & 0.08987 & 457.93 & 0.08830 & 577.85 & 0.09110 \\\\ \text{U-Net} & 489.18 & 0.05750 & 489.18 & 0.05634 & 588.68 & 0.05896 \\\\ \text{PCNO} & 349.10 & 0.07809 & 349.10 & 0.07824 & 451.88 & 0.07704 \\\\ \text{clawFNO} & 327.71 & 0.04532 & 327.71 & 0.04721 & 419.23 & 0.04586 \\\\ \text{w/ post-projection (CG)} & \text{-} & 2.93838 & \text{-} & 0.36412 & \text{-} & 2.92871 \\\\ \text{w/ post-projection (FFT-PCG)} & \text{-} & 0.09854 & \text{-} & 0.09470 & \text{-} & 0.09782 \\\\ \text{Sculpt} & 314.9 & 0.0570 & 314.9 & 0.0608 & 396.0 & 0.0596 \\\\ \hline \end{array}$$ **Train** measures one complete training epoch, and **Inference** measures one autoregressive prediction step. Neural timings use an NVIDIA A800 GPU. PALM uses 16 MPI processes, and its entry measures the time required to advance the simulation by one step of frame. Regarding imperfect initial fields, Sculpt's nested potential guarantees that each predicted velocity increment satisfies both discrete constraints. Consequently, the rollout preserves the constraint residuals of the initial state rather than correcting existing violations. In our current setting, predictions start from ground-truth LES frames that satisfy these constraints up to numerical precision. For non-admissible initial states, a possible extension is to apply a compatible pressure projection before rollout, restoring admissibility without requiring repeated projection at subsequent prediction steps. We leave the evaluation of this extension to future work. **References** [1] Hestenes, M. R., and Stiefel, E. Methods of Conjugate Gradients for Solving Linear Systems. Journal of Research of the National Bureau of Standards, 49(6):409–436, 1952. [2] Storti, M. A., et al. A FFT preconditioning technique for the solution of incompressible flow on GPUs. Computers & Fluids, 74:44–57, 2013.
AuthorsOpenReview ↗

Author Response to Reviewer TScW (Q2, Q3, Q4)

Q2: Real-City Layouts, Nonperiodic Boundaries, and Other Discretizations

Response to Q2:

We have not yet evaluated Sculpt on survey-derived city layouts or domains with nonperiodic horizontal boundaries. Extending Sculpt to nonperiodic domains would require adapting the boundary construction to distinguish impermeable walls from open boundaries and accommodate compatible inlet/outlet conditions.

For curved boundaries, cut cells, and alternative discretizations, preserving the guarantees would require compatible discrete operators that retain the divergence–curl and curl–gradient identities, together with a boundary assembly adapted to the new geometric representation [3, 4]. This also requires consistent treatment of face areas, cell volumes, and normal-flux degrees of freedom. Developing and validating these extensions remains an important direction for future work, as discussed in the Limitations and Future Work section.


Q3: Stability across Training Seeds and Held-Out Sequences

Response to Q3:

We report the completed comparisons across three training runs under the unseen-layout and unseen-morphology settings. MSE values are in units of 10−310^{-3} and are reported as mean ± sample standard deviation across training runs.

Unseen layouts

\hline \text{Method} & \text{MSE@1 }\downarrow & \text{MSE@4 }\downarrow & \text{MSE@8 }\downarrow & \text{MSE@16 }\downarrow \\\\ \hline \text{Dil-ResNet} & 1.14 \pm 0.38 & 10.52 \pm 2.73 & 36.53 \pm 6.60 & 105.82 \pm 32.31 \\\\ \text{TFNO} & 14.82 \pm 6.01 & 30.70 \pm 6.10 & 56.85 \pm 12.61 & 109.66 \pm 32.28 \\\\ \text{AFNO} & 10.49 \pm 1.76 & 29.01 \pm 1.24 & 44.00 \pm 1.85 & 63.13 \pm 2.05 \\\\ \text{AViT} & 7.64 \pm 0.25 & 17.88 \pm 0.26 & 33.53 \pm 0.92 & 59.31 \pm 1.32 \\\\ \text{P3D} & 1.08 \pm 0.06 & 6.67 \pm 0.72 & 19.11 \pm 2.10 & 44.52 \pm 1.29 \\\\ \text{U-Net} & 0.54 \pm 0.07 & 3.00 \pm 0.26 & 8.47 \pm 0.51 & 24.40 \pm 1.33 \\\\ \text{PCNO} & 21.11 \pm 0.01 & 34.69 \pm 0.08 & 40.97 \pm 0.25 & 49.48 \pm 0.43 \\\\ \text{clawFNO} & 17.28 \pm 0.18 & 27.33 \pm 0.26 & 33.74 \pm 0.64 & 44.68 \pm 2.01 \\\\ \text{Sculpt} & 0.21 \pm 0.01 & 0.91 \pm 0.15 & 3.35 \pm 0.48 & 14.59 \pm 1.14 \\\\ \hline \end{array}$$ **Unseen morphologies** $$\begin{array}{l|cccc} \hline \text{Method} & \text{MSE@1 }\downarrow & \text{MSE@4 }\downarrow & \text{MSE@8 }\downarrow & \text{MSE@16 }\downarrow \\\\ \hline \text{Dil-ResNet} & 1.76 \pm 0.62 & 15.15 \pm 3.87 & 48.43 \pm 7.86 & 128.20 \pm 36.53 \\\\ \text{TFNO} & 19.10 \pm 7.71 & 41.24 \pm 7.48 & 73.37 \pm 16.26 & 125.24 \pm 33.84 \\\\ \text{AFNO} & 13.96 \pm 2.17 & 38.64 \pm 1.38 & 59.62 \pm 2.04 & 91.07 \pm 4.56 \\\\ \text{AViT} & 10.57 \pm 0.21 & 27.82 \pm 0.80 & 53.25 \pm 1.95 & 88.56 \pm 2.80 \\\\ \text{P3D} & 1.67 \pm 0.09 & 10.10 \pm 0.75 & 28.44 \pm 1.76 & 61.66 \pm 1.44 \\\\ \text{U-Net} & 0.82 \pm 0.10 & 4.91 \pm 0.57 & 13.86 \pm 1.44 & 37.28 \pm 3.52 \\\\ \text{PCNO} & 25.80 \pm 0.01 & 44.61 \pm 0.25 & 54.52 \pm 0.60 & 68.09 \pm 1.08 \\\\ \text{clawFNO} & 21.40 \pm 0.16 & 36.39 \pm 0.14 & 47.06 \pm 0.70 & 64.31 \pm 3.10 \\\\ \text{Sculpt} & 0.30 \pm 0.02 & 1.62 \pm 0.23 & 6.00 \pm 0.60 & 24.34 \pm 0.78 \\\\ \hline \end{array}$$ We will report the complete multi-seed experimental results in the revised manuscript. ------ **Q4: Computational Cost Relative to U-Net, P3D, and Pressure Projection** **Response to Q4:** We have added **Appendix D.4** to the revised manuscript, including complexity analysis and runtime comparisons. The corresponding training cost, inference latency, peak memory, and parameter count comparisons are shown below:
Reviewer_qzgGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies 3D urban wind prediction while enforcing local mass conservation and wall impermeability. The authors propose Sculpt, a nested potential framework that combines a 3D U-Net for volume potential prediction and a surface graph network for boundary potential prediction. A discrete curl operator is then used to generate velocity updates that satisfy both constraints without additional pressure projection. The authors also introduce multiresolution learning to improve the prediction of large-scale flow structures. Experiments on the proposed UrbanWindFlow dataset, containing 85 PALM-LES sequences, show improved prediction accuracy across three generalization settings while maintaining near-zero physical constraint violations.

Critical strengths

S1. Meaningful Research Problem.

The paper addresses an important challenge in urban wind prediction by explicitly incorporating physical constraints into neural surrogate modeling. Ensuring mass conservation and wall impermeability is essential for producing physically consistent flow predictions.

S2. Principled Physics-Constrained Formulation.

The proposed nested potential formulation provides a mathematically grounded approach to enforcing mass conservation and wall impermeability by construction, avoiding the need for additional pressure projection during autoregressive prediction.

S3. Evaluation Across Multiple Generalization Settings.

The paper evaluates Sculpt under three distinct settings, including unseen building layouts, urban morphologies, and inflow conditions. These experiments provide useful evidence of the model's ability to generalize across different configurations within the proposed simulation benchmark.

Critical weaknesses

W1. Limited Methodological Novelty

The main method combines established mathematical operators with existing neural architectures, including 3D U-Net, GNN, and multiresolution fusion. Although the combination is reasonable, the paper does not clearly explain what is fundamentally new compared with existing physics-constrained approaches.

W2. Missing Strong and Relevant Baselines

Most baselines are established neural architectures rather than recent physics-constrained methods. More importantly, the paper does not compare against a model trained end-to-end with differentiable pressure projection or a simpler potential-based model without the surface GNN. These comparisons are important for demonstrating the actual advantage of Sculpt.

W3. Unclear Physical Validity of the Simulation Data

The PALM simulation setup raises two concerns. First, the paper appears to set the Coriolis parameter to zero while using geostrophic wind forcing, leaving the mechanism sustaining the flow unclear. Second, the simulation domain is only 48 m high, while the tallest buildings reach 36 m, potentially introducing strong boundary effects. The authors should clarify these settings and justify the physical reliability of the dataset.

W4. Limited Experimental Validation and Real-World Applicability

The evaluation is limited to short 16-step rollouts on a single synthetic LES dataset. Although three generalization settings are considered, the model has not been validated against real-world urban wind observations. Comparison with available measurements from urban meteorological stations or field campaigns would help assess its practical applicability. In addition, longer rollouts, multiple training seeds, and runtime comparisons are needed to better establish predictive robustness and computational efficiency.

Questions for authors

Q1. Methodological Novelty.

What are the key methodological contributions of Sculpt compared with existing physics-constrained neural models, particularly those based on potential representations?

Q2. Baseline Comparisons.

How does Sculpt compare with an end-to-end differentiable projection-based model or a simpler constraint-preserving potential model without the surface GNN? Can the authors clarify the necessity of the proposed architecture?

Q3. Simulation Setup.

Could the authors clarify how the flow is continuously driven in the PALM simulations when the Coriolis parameter is set to zero? Additionally, how is the relatively low domain height justified given the building heights?

Q4. Experimental Validation. Could the authors provide additional evidence of Sculpt's robustness and practical applicability, including longer prediction horizons, multiple training seeds, comparisons with real-world urban wind observations where feasible, and inference-time benchmarks?

5 comments
AuthorsOpenReview ↗

Author Response to Reviewer qzgG (W1 & Q1)

We thank the reviewer for the comments. Please find our responses to your concerns below.

W1 & Q1: Limited Methodological Novelty

Response to W1 & Q1:

We introduce a new formulation of urban wind prediction and a new way to solve it, supported by theoretical analysis and a dedicated dataset. Specifically:

  1. Problem formulation. We study how to jointly guarantee mass conservation and wall impermeability during autoregressive prediction across building layouts. Dense and heterogeneous urban geometries make these coupled constraints particularly challenging.

  2. Model design. We encode both constraints within a single nested volume–boundary potential representation, so the predicted updates satisfy them by construction without a pressure solve at each step.

  3. Finding and theoretical analysis. We identify that backpropagation through the curl attenuates large-scale gradients relative to small-scale signals. We explain this imbalance through spectral analysis and design multiresolution potential readouts to address it.

  4. Dataset. We introduce UrbanWindFlow, a building-resolving LES dataset of three-dimensional instantaneous urban wind fields, to evaluate prediction accuracy, physical constraint satisfaction, and generalization across layouts, morphology classes, and inflow conditions.

We have explained how our formulation differs from existing approaches [1–3] in Section 1 (Lines 054–088), Section 2.2 (Lines 191–208), and Appendix E (Lines 1247–1258). Our architectural choices address specific needs of this formulation. The multiresolution readouts give large-scale flow structures separate prediction branches (Section 3.2, Lines 283–304; Appendix A.2). The surface GNN propagates information along the boundary mesh, where connectivity matters more than Euclidean proximity (Section 3.3, Lines 307–311).

We maintain that we do not simply combine or stack existing modules. We design them around the structure and learning properties of the two coupled constraints. These components each address a specific difficulty in learning potential representations for complex building geometries. We will revise the manuscript to better emphasize these methodological contributions and distinguish our work from existing approaches.

References

[1] Li X, et al. Project and Generate: Divergence-Free Neural Operators for Incompressible Flows. arXiv:2603.24500, 2026.

[2] Huang Y, Greenberg D S. Geometric and Physical Constraints Synergistically Enhance Neural PDE Surrogates. arXiv:2506.05513, 2025.

[3] Wandel N, Weinmann M, Klein R. Teaching the incompressible Navier–Stokes equations to fast neural surrogate models in three dimensions. Physics of Fluids, 2021.

AuthorsOpenReview ↗

Author Response to Reviewer qzgG (W2 & Q2)

W2 & Q2: Missing Strong and Relevant Baselines

Response to W2 & Q2:

Thank you for the suggestion. Our main comparison already includes POD, whose snapshot-derived basis inherits the divergence-free structure, and DMD & Time-Mean, which preserve both mass conservation and wall impermeability through linear combinations of historical fields from the same layout.

To further strengthen the comparison, we have added two recent physics-constrained neural baselines: PCNO [1], trained end-to-end with differentiable pressure projection and loss computed on the projected output, and clawFNO [2], a simpler potential-based model without a surface GNN. Their original implementations do not directly support our new setting with coupled constraints across heterogeneous building layouts, so we adapted them to our three-dimensional staggered grid and building boundaries. For clawFNO, we use fixed boundary-potential conditions to preserve both constraints. The table below presents the results across all three settings.

Unseen layouts

\hline \text{Method} & \begin{array}{c}\text{MSE}\\\\\text{@1 }\downarrow\end{array} & \begin{array}{c}\text{MSE}\\\\\text{@4 }\downarrow\end{array} & \begin{array}{c}\text{MSE}\\\\\text{@8 }\downarrow\end{array} & \begin{array}{c}\text{MSE}\\\\\text{@16 }\downarrow\end{array} & \begin{array}{c}\text{Divergence}\\\\\text{@16 }\downarrow\end{array} & \begin{array}{c}\text{Wall penetration}\\\\\text{@16 }\downarrow\end{array} \\\\ \hline \text{PCNO} & 21.11 \pm 0.01 & 34.69 \pm 0.08 & 40.97 \pm 0.25 & 49.48 \pm 0.43 & < 10^{-6} & 0 \\\\ \text{clawFNO} & 17.28 \pm 0.18 & 27.33 \pm 0.26 & 33.74 \pm 0.64 & 44.68 \pm 2.01 & < 10^{-6} & 0 \\\\ \text{Sculpt} & \mathbf{0.21 \pm 0.01} & \mathbf{0.91 \pm 0.15} & \mathbf{3.35 \pm 0.48} & \mathbf{14.59 \pm 1.14} & < 10^{-6} & < 10^{-6} \\\\ \hline \end{array}$$ **Unseen morphologies** $$\begin{array}{l|cccc|cc} \hline \text{Method} & \begin{array}{c}\text{MSE}\\\\\text{@1 }\downarrow\end{array} & \begin{array}{c}\text{MSE}\\\\\text{@4 }\downarrow\end{array} & \begin{array}{c}\text{MSE}\\\\\text{@8 }\downarrow\end{array} & \begin{array}{c}\text{MSE}\\\\\text{@16 }\downarrow\end{array} & \begin{array}{c}\text{Divergence}\\\\\text{@16 }\downarrow\end{array} & \begin{array}{c}\text{Wall penetration}\\\\\text{@16 }\downarrow\end{array} \\\\ \hline \text{PCNO} & 25.80 \pm 0.01 & 44.61 \pm 0.25 & 54.52 \pm 0.60 & 68.09 \pm 1.08 & < 10^{-6} & 0 \\\\ \text{clawFNO} & 21.40 \pm 0.16 & 36.39 \pm 0.14 & 47.06 \pm 0.70 & 64.31 \pm 3.10 & < 10^{-5} & 0 \\\\ \text{Sculpt} & \mathbf{0.30 \pm 0.02} & \mathbf{1.62 \pm 0.23} & \mathbf{6.00 \pm 0.60} & \mathbf{24.34 \pm 0.78} & < 10^{-6} & < 10^{-6} \\\\ \hline \end{array}$$ **Unseen inflows** $$\begin{array}{l|cccc|cc} \hline \text{Method} & \begin{array}{c}\text{MSE}\\\\\text{@1 }\downarrow\end{array} & \begin{array}{c}\text{MSE}\\\\\text{@4 }\downarrow\end{array} & \begin{array}{c}\text{MSE}\\\\\text{@8 }\downarrow\end{array} & \begin{array}{c}\text{MSE}\\\\\text{@16 }\downarrow\end{array} & \begin{array}{c}\text{Divergence}\\\\\text{@16 }\downarrow\end{array} & \begin{array}{c}\text{Wall penetration}\\\\\text{@16 }\downarrow\end{array} \\\\ \hline \text{PCNO} & 33.60 \pm 14.65 & 70.81 \pm 18.24 & 81.93 \pm 20.34 & 87.48 \pm 23.16 & < 10^{-6} & 0 \\\\ \text{clawFNO} & 30.69 \pm 12.60 & 58.00 \pm 15.29 & 66.26 \pm 16.85 & 70.32 \pm 18.67 & < 10^{-6} & 0 \\\\ \text{Sculpt} & \mathbf{0.89 \pm 0.67} & \mathbf{6.35 \pm 5.41} & \mathbf{19.53 \pm 15.00} & \mathbf{46.31 \pm 25.52} & < 10^{-6} & < 10^{-6} \\\\ \hline \end{array}$$ MSE values are reported in units of $10^{-3}$. The entries show mean ± sample standard deviation across three training seeds. Divergence and wall penetration denote the maximum absolute constraint residuals at the 16-step horizon. We achieve substantially lower prediction errors than both newly added baselines across all settings. We will report the mean and standard deviation across independent training seeds for all models in the revised manuscript. We will also report longer-horizon results for all models, including the newly added baselines, in Appendix D.2. **References** [1] Xu Q, et al. Physically consistent and uncertainty-aware learning of spatiotemporal dynamics. arXiv:2510.21023, 2025. [2] Liu N, et al. Harnessing the Power of Neural Operators with Automatically Encoded Conservation Laws. ICML, 2024.
AuthorsOpenReview ↗

Author Response to Reviewer qzgG (W3 & Q3, W4 & Q4)

W3 & Q3: Unclear Physical Validity of the Simulation Data

Response to W3 & Q3:

Thank you for raising these two concerns about the simulation setup.

Regarding the mechanism sustaining the flow, our description in the Appendix B.1 (Line 875) was imprecise. We adopted the term “geostrophic” from PALM’s parameter names without clearly distinguishing initialization from continuous forcing. With the Coriolis force disabled, ug_surface and vg_surface specify the initial horizontal velocity profiles, as explained on the PALM tutorial [https://docs.palm-model.org/25.10/Reference/LES_Model/Namelists/#conserve_volume_flow]. The flow is sustained by volume-flow feedback (conserve_volume_flow = .T.), with the target flow determined from the initial profiles. This is an engineering configuration documented by PALM [https://docs.palm-model.org/25.10/Reference/LES_Model/Namelists/#conserve_volume_flow]. This configuration has also been adopted in previous PALM-based studies of street-canyon and urban flows, which likewise conserve volume flow while neglecting the Coriolis force [1,2]. We revised the manuscript to accurately describe this mechanism.

Regarding boundary effects, these are a general consideration in finite-domain simulations. Previous LES studies of periodic street canyons have used domain-height-to-building-height ratios ranging from 2 to 6 [3–5]. In UrbanWindFlow, individual roof heights range from 6 to 40 m within a 48 m-high domain, corresponding to domain-height-to-individual-roof-height ratios of 1.2–8.0. Across the dataset, over 85% of the total building footprint area corresponds to roof heights at or below 24 m, yielding ratios of at least 2. Our dataset design considers the available overhead space by including layouts with both limited and larger clearances between the tallest buildings and the domain top. We clarify these details in the revised manuscript.

References

[1] Keskinen J P, Hellsten A. Ensembles in Urban Large Eddy Simulations with Changing Wind Direction. Boundary-Layer Meteorology, 2025.

[2] Rienecker H, et al. Energy optimal 3D flight path planning for unmanned aerial vehicle in urban environments. CEAS Aeronautical Journal, 2023.

[3] Li, et al. Large-eddy simulation of flow and pollutant transport in urban street canyons with ground heating. Boundary-layer meteorology, 2010.

[4] Li, et al. Flow and pollutant transport in urban street canyons of different aspect ratios with ground heating. Boundary-layer meteorology, 2012.

[5] Cheng W C, Liu C H. Large-Eddy Simulation of Flow and Pollutant Transports in and Above Two-Dimensional Idealized Street Canyons. Boundary-Layer Meteorology, 2011.


W4 & Q4: Limited Experimental Validation and Real-World Applicability

Response to W4 & Q4:

Real-world applicability.

We agree that real-world observations would help assess practical applicability, but our methodological contribution is already supported by theory and LES experiments. First, our primary contribution is a neural-network design that inherently satisfies coupled constraints. We have validated the soundness and effectiveness of this approach through extensive experiments and theoretical analysis. Sparse and noisy urban observations can provide valuable pointwise validation, but do not by themselves supply the dense, time-resolved three-dimensional reference fields needed to jointly evaluate full-field prediction accuracy and constraint violations. We use building-resolving LES to support this controlled assessment.

Second, real-world observations typically consist of incomplete measurements at a limited number of locations and contain measurement errors and noise. These limited measurements may not necessarily provide more effective validation. We acknowledge that validation against real-world observations can assess the practical applicability of our method. However, this direction also involves other complex factors, such as the observation-based initialization, uncertainty in geometry and forcing, and the simulation-to-reality gap. Overall, it appears to extend beyond the scope of the present paper. We will clarify this scope and limitation in the revised manuscript and identify validation against urban field measurements as an important direction for future work.

Experimental Validation

We have added three sets of experiments: longer prediction horizons, comparisons across multiple training seeds, and runtime cost. We report their results separately below.

Reviewer_NZtxOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work predicts transient urban winds using a neural surrogate while guaranteeing local mass conservation and respecting wall boundaries. The work uses a nested potential framework that build the geometry dependent constraints into the parametrization, and the volume potential is parameterized at multiple resolutions. This results in divergence-free velocity updates. In addition the work provides a benchmakr dataset for evaluation. The proposed method significantly improves upon comprehensive baselines.

Critical strengths
  1. This work enforce kinematic admissibility by construction which is a solid approach.

  2. Good results significantly improving prediction accuracy compared with comprehensive baselines and evaluation settings.

  3. Data and code are available.

Critical weaknesses
  1. Ablations could also include learning the boundary scalar potential.

  2. Missing comparison of computation, memory, and latency numbers, specifically to an optimized projection solver.

  3. This is an applications paper in a very specific field.

Questions for authors

Why 16 steps? What about longer trajectories?

3 comments
AuthorsOpenReview ↗

Author Response to Reviewer NZtx (Weakness 1, 2, 3)

We thank the reviewer for the comments. Please find our responses to your concerns below.

W1: Ablations could also include learning the boundary scalar potential.

Response to W1:

You can find the existing w/o boundary GG variant ablates the learned boundary scalar potential module, with results reported in the ablation study (Tables 4, 5, 8).


W2: Missing comparison of computation, memory, and latency numbers, specifically to an optimized projection solver.

Response to W2:

We have added a section Appendix D.4 to the revised manuscript, including complexity analysis and runtime comparisons. We provide the comparison including an optimized pressure-projection solver. The pressure-projection baseline corresponds to our “w/ post-projection” ablation variant. We compare the original Conjugate Gradient (CG) solver [1] with our GPU implementation of a Fast Fourier Transform–preconditioned Conjugate Gradient (FFT-PCG) solver [2]. Both implementations use the same trained direct predictor, discrete pressure equation, building boundary conditions, and solver tolerance.

\hline \text{Method} & \begin{array}{c}\text{Unseen layouts: }\\\\\text{Train (s/epoch)}\end{array} & \begin{array}{c}\text{Unseen layouts: }\\\\\text{Inference (s/step)}\end{array} & \begin{array}{c}\text{Unseen morphologies: }\\\\\text{Train (s/epoch)}\end{array} & \begin{array}{c}\text{Unseen morphologies: }\\\\\text{Inference (s/step)}\end{array} & \begin{array}{c}\text{Unseen inflows: }\\\\\text{Train (s/epoch)}\end{array} & \begin{array}{c}\text{Unseen inflows: }\\\\\text{Inference (s/step)}\end{array} \\\\ \hline \text{PALM (LES; 16 MPI processes)} & \text{—} & 4.32 & \text{—} & 4.32 & \text{—} & 4.32 \\\\ \text{Dil-ResNet} & 457.93 & 0.08987 & 457.93 & 0.08830 & 577.85 & 0.09110 \\\\ \text{U-Net} & 489.18 & 0.05750 & 489.18 & 0.05634 & 588.68 & 0.05896 \\\\ \text{PCNO} & 349.10 & 0.07809 & 349.10 & 0.07824 & 451.88 & 0.07704 \\\\ \text{clawFNO} & 327.71 & 0.04532 & 327.71 & 0.04721 & 419.23 & 0.04586 \\\\ \text{w/ post-projection (CG)} & \text{-} & 2.93838 & \text{-} & 0.36412 & \text{-} & 2.92871 \\\\ \text{w/ post-projection (FFT-PCG)} & \text{-} & 0.09854 & \text{-} & 0.09470 & \text{-} & 0.09782 \\\\ \text{Sculpt} & 314.9 & 0.0570 & 314.9 & 0.0608 & 396.0 & 0.0596 \\\\ \hline \end{array}$$ **Train** measures one complete training epoch, and **Inference** measures one autoregressive prediction step. Neural timings use an NVIDIA A800 GPU. PALM uses 16 MPI processes, and its entry measures the time required to advance the simulation by one step of frame. **References** [1] Methods of Conjugate Gradients for Solving Linear Systems. Journal of Research of the National Bureau of Standards, 1952. [2] A FFT preconditioning technique for the solution of incompressible flow on GPUs. Computers & Fluids, 2013. ------- **W3: This is an applications paper in a very specific field.** **Response to W3:** Urban wind prediction and related physical modeling have long been of interest to the machine-learning community [1–6]. We would like to clarify that our main contribution is a new neural-network design that jointly enforces mass conservation and wall impermeability for transient urban wind prediction. This problem is particularly important in urban wind prediction because dense and heterogeneous building geometries introduce extensive boundary constraints that must remain compatible with mass conservation. We address this challenge through a nested potential representation that satisfies both constraints by construction. We further characterize the gradient attenuation induced by the curl operator and introduce multiresolution readouts accordingly, while learning boundary potentials through information propagation along the boundary mesh. We will revise the manuscript to make clear that the novelty lies in the methodology of physics-constrained learning, rather than in the application scenario itself. **References** [1] Giral, F., et al. GenDA: Generative Data Assimilation on Complex Urban Areas via Classifier-Free Diffusion Guidance. ICML, 2026. [2] Xin, W., et al. UrbanGraph: Physics-Informed Spatio-Temporal Dynamic Heterogeneous Graphs for Urban Microclimate Prediction. ICLR, 2026. [3] Hettige, K. H., et al. AirPhyNet: Harnessing Physics-Guided Neural Networks for Air Quality Prediction. ICLR, 2024. [4] Pfaff, T., et al. Learning Mesh-Based Simulation with Graph Networks. ICLR, 2021. [5] Saad, N., et al. Guiding Continuous Operator Learning through Physics-based Boundary Constraints. ICLR, 2023. [6] Wang, H., et al. Neural Fluid Simulation on Geometric Surfaces. ICLR, 2025.
AuthorsOpenReview ↗

Author Response to Reviewer NZtx (Question 1)

Q1. Why 16 steps? What about longer trajectories?

Response to Q1:

Our autoregressive evaluation protocol follows precedents in prior studies, including DynaBench [1], FactFormer [2], and P3D [3], which include 16-step predictions. We report performance at 1, 4, 8, and 16 steps, already revealing clear differences between short- and longer-horizon prediction. To further address the reviewer’s concern, we have added results for all models at 24 and 32 steps.

Unseen layouts:

\hline \text{Method} & \begin{array}{c}\text{MSE}\\\\\text{@24 }\downarrow\end{array} & \begin{array}{c}\text{MSE}\\\\\text{@32 }\downarrow\end{array} & \begin{array}{c}\text{Divergence}\\\\\text{@24 }\downarrow\end{array} & \begin{array}{c}\text{Wall penetration}\\\\\text{@24 }\downarrow\end{array} & \begin{array}{c}\text{Divergence}\\\\\text{@32 }\downarrow\end{array} & \begin{array}{c}\text{Wall penetration}\\\\\text{@32 }\downarrow\end{array} \\\\ \hline \text{DMD} & 62.02 & 59.81 & < 10^{-6} & 0 & < 10^{-6} & 0 \\\\ \text{Time-Mean} & 50.33 & 48.66 & < 10^{-6} & 0 & < 10^{-6} & 0 \\\\ \text{POD} & 56.34 & 56.13 & < 10^{-5} & 1.61 & < 10^{-5} & 1.60 \\\\ \text{Dil-ResNet} & 99.92 & 128.81 & 0.57 & 1.35 & 0.47 & 1.43 \\\\ \text{TFNO} & 124.09 & 147.44 & 0.83 & 1.52 & 0.55 & 1.50 \\\\ \text{AFNO} & 85.30 & 499.08 & 5.75 & 4.24 & 108.51 & 50.59 \\\\ \text{AViT} & 72.08 & 80.27 & 1.93 & 1.73 & 1.89 & 1.59 \\\\ \text{P3D} & 61.44 & 1957.13 & 0.85 & 1.02 & 154.27 & 35.02 \\\\ \text{U-Net} & 41.96 & 53.64 & 0.14 & 1.02 & 0.14 & 1.19 \\\\ \text{PCNO} & 54.81 & 60.74 & < 10^{-6} & 0 & < 10^{-6} & 0 \\\\ \text{clawFNO} & 55.31 & 79.77 & < 10^{-5} & 0 & < 10^{-5} & 0 \\\\ \text{Sculpt} & \mathbf{29.64} & \mathbf{42.02} & < 10^{-5} & < 10^{-6} & < 10^{-5} & < 10^{-6} \\\\ \hline \end{array}$$ Unseen morphologies: $$\begin{array}{l|cc|cc|cc} \hline \text{Method} & \begin{array}{c}\text{MSE}\\\\\text{@24 }\downarrow\end{array} & \begin{array}{c}\text{MSE}\\\\\text{@32 }\downarrow\end{array} & \begin{array}{c}\text{Divergence}\\\\\text{@24 }\downarrow\end{array} & \begin{array}{c}\text{Wall penetration}\\\\\text{@24 }\downarrow\end{array} & \begin{array}{c}\text{Divergence}\\\\\text{@32 }\downarrow\end{array} & \begin{array}{c}\text{Wall penetration}\\\\\text{@32 }\downarrow\end{array} \\\\ \hline \text{DMD} & 79.03 & 75.53 & < 10^{-6} & 0 & < 10^{-6} & 0 \\\\ \text{Time-Mean} & 64.02 & 61.57 & < 10^{-6} & 0 & < 10^{-6} & 0 \\\\ \text{POD} & 83.81 & 82.40 & < 10^{-5} & 1.76 & < 10^{-5} & 1.72 \\\\ \text{Dil-ResNet} & 118.18 & 142.58 & 0.50 & 1.30 & 0.54 & 1.63 \\\\ \text{TFNO} & 122.07 & 130.52 & 1.06 & 2.43 & 1.07 & 2.35 \\\\ \text{AFNO} & 128.88 & 1024.30 & 54.99 & 7.23 & 306.60 & 56.91 \\\\ \text{AViT} & 100.34 & 107.63 & 2.09 & 1.84 & 2.09 & 1.80 \\\\ \text{P3D} & 1501.85 & 9802.59 & 215.39 & 89.13 & 196.83 & 141.94 \\\\ \text{U-Net} & 60.25 & 71.83 & 0.19 & 1.20 & 0.19 & 1.11 \\\\ \text{PCNO} & 78.92 & 87.10 & < 10^{-6} & 0 & < 10^{-6} & 0 \\\\ \text{clawFNO} & 77.44 & 103.82 & < 10^{-5} & 0 & < 10^{-5} & 0 \\\\ \text{Sculpt} & \mathbf{43.98} & \mathbf{58.34} & < 10^{-5} & < 10^{-6} & < 10^{-5} & < 10^{-6} \\\\ \hline \end{array}$$ Sculpt achieves the lowest MSE in extended-horizon comparisons, while maintaining divergence below $10^{-5}$ and wall penetration below $10^{-6}$. We also observe that Time-Mean and POD outperform several autoregressive neural baselines at longer horizons. A plausible explanation is that, under fixed geometry and prescribed forcing, smooth or low-dimensional predictions can remain close to the persistent background flow, while autoregressive neural models accumulate prediction errors. Furthermore, the long-term prediction stability of neural surrogate models is an important issue. We thank the reviewer for this insight, and addressing this challenge will be a key objective of our future research. **References** [1] Dulny, A., Hotho, A., and Krause, A. DynaBench: A benchmark dataset for learning dynamical systems from low-resolution data. ECML PKDD, 2023. [2] Li, Z., Shu, D., and Farimani, A. B. Scalable Transformer for PDE Surrogate Modeling. NeurIPS, 2023. [3] Holzschuh, B., Kohl, G., Redinger, F., and Thuerey, N. P3D: Scalable Neural Surrogates for High-Resolution 3D Physics Simulations with Global Context. arXiv:2509.10186, 2025.
Area_Chair_ATbrOpenReview ↗

Reviewer NZtx, please respond to the authors' comments

Reviewer NZtx, please respond to the authors' comments. Thank you.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.