acceptodds
Withdrawn from ICLR 2027

Coarse-to-fine Physics-guided Mixture-of-experts for Urban Wind Field Reconstruction from a Single Radar Column

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Accurate low-altitude urban wind fields are essential for UAV safety, pollutant dispersion assessment, and wind analysis around buildings. However, reconstructing high-resolution 3D urban wind fields from a single vertical radar column is highly challenging due to extremely sparse observations. Existing methods often rely on distributed sensors or multiple LiDARs, while direct sparse-to-dense mapping entangles large-scale wind field completion with fine-scale structure recovery, and shared representations struggle to capture diverse local flow patterns. To address these challenges, we propose a coarse-to-fine urban wind field reconstruction framework that decomposes the task into low-resolution wind field completion and multi-stage super-resolution reconstruction, progressively bridging the large spatial gap between sparse observations and dense targets. We further introduce a physics-guided mixture-of-experts module with different physical priors that encourages different experts to specialize in distinct local flow patterns, including general air, building boundaries, vortices, highspeed regions, and ordinary areas. To mitigate the mismatch between ideal and predicted intermediate states during cascaded reconstruction, we develop a momentum-energy-guided state selection mechanism that prioritizes representative predicted states according to their momentum and kineticenergy discrepancies. Our experiments achieve 0.1408 MAE and 0.0404 MSE, reducing these errors by 46.00% and 70.19% over the state-of-theart method, respectively, demonstrating the effectiveness of our proposed method.Code is available at https://anonymous.4open.science/r/windreconstruction-anonymous-4C75/

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
3D wind field reconstruction, sparse observation, super-resolution, mixture-of-experts, physics-informed learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_6f8uOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper reconstructs high-resolution 3D urban wind fields from a single vertical LiDAR column, together with building geometry and inflow conditions. It proposes a coarse-to-fine framework consisting of low-resolution completion and two super-resolution stages, a physics-guided mixture-of-experts module with region-specific experts and losses (general air, building boundary, vortex, high-speed, and ordinary regions), and a momentum–energy-guided mechanism that decides whether the second super-resolution stage is trained on ground-truth or predicted intermediate states. Experiments on 4,824 samples report an MAE of 0.1408, a 46% reduction over the best baseline.

Critical strengths
  1. The introduction is well written: the two challenges are clearly explained and directly correspond to the proposed components.
  2. The comparison includes a large number of baselines, and the ablations analyze each module as well as individual experts.
Critical weaknesses
  1. Gains do not come from the physics-guided designs. The largest improvement comes from the sampling mechanism (Table 2), but simply training on predicted inputs already achieves most of it (Pred-only 0.1485 vs. ours 0.1408, Table 5); this is essentially scheduled sampling, which is not cited. The direct reconstruction baseline already outperforms all compared methods, and several learned baselines are worse than nearest-neighbor interpolation, raising concerns about baseline adaptation.
  2. Role of the LiDAR/Radar input and generalization are not validated. Since building geometry and inflow conditions are given as input, an ablation without the LiDAR column is needed to show its contribution. The data source, sample variation, and split protocol are not explained, so generalization to unseen conditions is unclear.Unclear writing and task definition.
  3. Presentation and reproducibility. "Radar" and "LiDAR" are used interchangeably, and the sampling mechanism has several different names. The repository only contains checkpoint files, which appear incomplete, without model or training code.
Questions for authors
  1. What is the performance without the LiDAR/Radar input, using only building geometry and inflow conditions?
  2. What is the data source, what varies among samples, and does the test set contain unseen geometries or inflow conditions?
  3. Beyond the Pred-only setting, what is the specific contribution of the momentum–energy schedule, and how does it differ from scheduled sampling?
Reviewer_aNLcOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes a 3D urban wind reconstruction method from sparse LiDAR observations using a physics-informed mixture of experts model, where each expert specializes in a type of flow structure that can be seen in the urban environment. This is combined with a coarse-to-fine reconstruction structure to go from sparse observations to fine targets.

Critical strengths

S1. The idea of using a combination of a standard mixture of experts and a physics-based mixture of experts separated by types of flow structures seems interesting, as incorporating domain knowledge into the model structure may help.

Critical weaknesses

W1. I see a clear lack of generalization capabilities. As very little information about the dataset used is given, I am guessing that the 4824 samples and 24 inflow directions come from 24 trajectories of 201 time steps (as no other flow conditions or geometry are said to be changing). Nor are the 24 inflow directions specified. That being said, several samples may be correlated or change little, and coming from the same geometry and flow conditions means there is not enough generalization being shown. Additionally, there is no clear explanation of how the splits are done, whether samples are randomly selected, split by inflow directions, etc.

W2. The gap between the proposed model and all the baselines seems unreasonable. The big gap makes me think there may be a problem either with the implementation, data processing, or another part of the pipeline. Considering that for MAE the best baseline achieves 0.26, and even the worst-case ablation of the proposed method (direct reconstruction, without MoE, coarse-to-fine, or physical feedback) achieves an MAE of 0.22, this makes me think there may be some unresolved issues with the baseline results. The authors should clarify how the baselines were adapted to the task, data pipeline, hyperparameter tuning, computational budget, etc.

Questions for authors

Q1. In Figures 1.b, 1.c, and 3, I observe that there is no fluid field between the buildings, but only outside the urban area (this may be because of the z=2 cut, but the same is shown in the 3D figure in 1.b). If that is the case, why is that? Is the flow between buildings not the real area of interest?

Q2. Why are only absolute metrics reported? I consider that relative metrics such as relative L2 and L1 errors might be more informative for fluid flow reconstruction.

Q3. You define your method as physics-informed; however, no physics metric is reported during evaluation to actually show that your method is maintaining physical constraints better, such as divergence, vorticity, boundary condition adherence, etc.

Q4. Dataset information should be better reported and explained. How were the simulations computed? What method was used? What were the conditions of the simulations? What conditions are being varied for the different samples? How are the splits done, etc.?

Q5. For the physics-based MoE, I would like to see what the physics-based approach contributes versus the increase in model size and capacity. For example, a comparison with a conventional MoE only, matching the architecture, number of experts, and computational budget, would be useful. Also, an analysis of how well the router captures the different flow-type regions would be nice to have.

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

This paper studies the reconstruction of high-resolution 3D urban wind fields from extremely sparse observations along a single vertical sensing column. To address the large spatial gap between sparse observations and dense wind-field targets, the authors propose a coarse-to-fine reconstruction framework that progressively performs low-resolution field completion and two-stage super-resolution. The framework further incorporates a physics-guided mixture-of-experts (MoE) module, which employs region-specific physical priors to model heterogeneous flow structures, including building boundaries, vortices, and high-speed regions. In addition, a momentum–energy-guided state selection mechanism is introduced to mitigate the mismatch between ground-truth and predicted intermediate states during cascaded reconstruction. Experiments on 4,824 wind-field samples covering 24 inflow directions show that the proposed method achieves an MAE of 0.1408 and an MSE of 0.0404, corresponding to reported reductions of 46.00% and 70.19%, respectively, compared with the state-of-the-art baseline. Ablation studies are also conducted to evaluate the contributions of the proposed framework and its key components.

Critical strengths
  1. Physics-guided modeling of heterogeneous flow regions. The proposed mixture-of-experts module incorporates different physical priors for distinct flow regions, including building boundaries, vortices, and high-speed regions. This provides an interpretable motivation for expert specialization and allows the model to account for the heterogeneous flow patterns encountered in urban environments.

  2. A targeted mechanism for intermediate-state mismatch. The momentum–energy-guided state selection mechanism explicitly addresses the discrepancy between ground-truth and predicted intermediate states in cascaded reconstruction. The paper also includes ablation studies comparing different reconstruction configurations, expert components, and state-selection strategies, providing an initial assessment of the contributions of the proposed components.

Critical weaknesses
  1. In Figure 2, some labels and annotations are relatively small, and the repeated textual information makes the figure slightly cluttered. Increasing the font size and removing some redundant text would improve the figure's readability and overall presentation.

  2. The quantitative evaluation relies primarily on MAE, MSE, RMSE, and PSNR, which provide limited complementary information because RMSE is directly derived from MSE and PSNR is mathematically determined by MSE when the dynamic range is fixed. Although SSIM is reported in Table 5, it is not included in the main comparison with existing methods (Table 1) or the overall framework ablation study (Table 2), limiting the consistency of structural evaluation across experiments. Moreover, given that the proposed method incorporates physics-guided constraints, the evaluation would be strengthened by directly assessing the physical consistency of the reconstructed wind fields. For example, reporting divergence residuals could help determine whether the physical constraints improve the consistency of the reconstructed fields. Additional component-wise error analysis could also provide a more detailed assessment of reconstruction performance.

Questions for authors
  1. Data Generation and Dataset Construction. Could the authors provide a more detailed description of the dataset generation process? Specifically, are the 4,824 wind-field samples obtained from CFD simulations or real-world measurements? How are the 3D ground-truth wind fields generated, and how are the sparse observations along a single sensing column derived from them? Additionally, how are the building geometries and inflow conditions configured, and how are the samples divided into training, validation, and test sets?

  2. Code Availability and Reproducibility. The provided anonymous repository includes a frozen model bundle and a final-stage inference script, but does not appear to provide the complete training implementation or the upstream components required to reproduce the full reconstruction pipeline. Could the authors clarify whether the complete source code will be made available, including the model architecture, loss functions, training configurations, and the implementation of the momentum–energy-guided state selection mechanism? In addition, providing the evaluation scripts and sufficient implementation details to reproduce the reported results would improve the reproducibility of the work.

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

Related papers

Discussion (0)

Sign in to comment.