acceptodds
Withdrawn from ICLR 2027

WindMiL: Equivariant Graph Learning for Building Wind Loads

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Wind loading prediction is safety-critical for building design, yet it is largely absent from ML benchmarks because high-fidelity aerodynamic labels are expensive to generate and defined on irregular building surfaces. Current practice relies on wind tunnel testing, design codes, and computational fluid dynamics with large-eddy simulation (LES), where each case requires more than 24 hours, making parametric exploration infeasible. Here we introduce WindMiL, a graph learning framework that advances both ML and wind engineering: it provides a realistic LES-based benchmark for studying physical inductive bias, extrapolation, and symmetry, while enabling LES-based surrogate prediction of pressure loads on buildings. We generate 462 LES simulations of low-rise buildings under realistic atmospheric boundary layer flow, spanning roof shapes and wind directions. The central challenge is not only accuracy, but also respecting the physical constraints: reflecting a building–wind configuration should reflect its predicted pressure field. However, training on reflected examples does not guarantee this consistency. WindMiL encodes this order-2 symmetry directly in the architecture, creating a testbed for equivariance versus data augmentation. We compare it with a pointwise MLP, GraphSAGE, and reflection-augmented GraphSAGE at different data splits. Across random, barycentric extrapolation, and edge-extrapolation splits, WindMiL achieves the highest overall accuracy for surface pressure coefficients while maintaining exact reflection symmetry. In the OOD edge extrapolation, WindMiL reduces mean RMSE by 7.9% relative to reflection-augmented GraphSAGE and improves hitrate relative to non-augmented GraphSAGE, by approximately 5%. These gains also translate to integrated drag and lift coefficients, with best hitrates and exact symmetry. These results show how a task-specific symmetry can improve generalization while enforcing a physical constraint that data augmentation only approximates. This study establishes wind loading prediction as a benchmark for geometric deep learning and ML-based computational wind engineering.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Geometric Deep Learning, Graph Neural Networks, Reflection Equivariance, Wind load prediction, Computational Fluid Dynamics, Large-eddy Simulations, Benchmark

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_qt9cOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

WindMiL introduces a dataset of 462 building-flow simulations covering 66 geometries and seven wind directions. It also proposes a graph model that combines representations of original and reflected geometries to enforce reflection symmetry. The experiments examine pressure and load prediction across random and geometry-extrapolation splits.

Critical strengths

The dataset is a useful contribution. I appreciate the effort involved in generating the LES cases and evaluating both mean pressure and pressure variability. The geometry-based splits also make the evaluation more informative by testing how well models handle shapes outside the training region.

The physical motivation is sensible, and the pressure results are encouraging. Reflection symmetry is appropriate for the setting considered here. WindMiL achieves the lowest reported mean-pressure RMSE across all three original test splits. On edge extrapolation, the improvement over augmented GraphSAGE is about 7.9% (Table 1).

Critical weaknesses

!) It is unclear how much the proposed averaging strategy adds beyond a simpler alternative. A straightforward comparison would average a baseline’s predictions on the original and reflected inputs. This would also enforce exact reflection consistency. Without that comparison, it is difficult to tell how much benefit comes specifically from averaging learned representations. Results across multiple training seeds would also help establish whether the accuracy improvements are consistent.

  1. I could not reconcile the descriptions of the load-prediction experiment. Section 4.2 suggests that Table 2 shows improved pressure predictions translating into better integrated loads. Appendix A.4, however, describes separately trained graph-level predictors. These support different conclusions: accurate force predictions from a separate model do not establish that integrating the predicted pressure field gives accurate forces. The construction of force-standard-deviation targets and the hitrate thresholds also need clarification.

  2. The authors have not benchmarked their approach against state-of-the-art neural solvers such as transolver/++, abupt, geotransolver, etc. I see this as the biggest limitation and therefore a reason for rejection.

Questions for authors
  1. How does WindMiL compare with averaging the original/reflected outputs of GS and GSDAG? How consistent are the gains across training seeds?

  2. Does Table 2 use integrated pressure predictions or separately trained load predictors? Are force-standard-deviation labels calculated from LES force time histories, and what numerical hitrate thresholds are used?

  3. Appendix A.1 lists 361 training/validation cases and 105 test cases, totaling 466 rather than 462. Could you clarify the correct split counts?

Reviewer_v8N7OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes WindMiL, a reflection-equivariant graph neural network to predict wind loads on buildings and integrated forces such as lift and drag, and compares the model to MLP and GraphSAGE baselines. The authors also introduce a dataset consisting of 462 large-eddy simulations varying across 66 geometries and 7 wind inflow angles.

Critical strengths

S1. The paper addresses a scientifically relevant problem for wind engineering through the use of machine learning.

S2. The LES dataset represents a computational effort and could be a good resource for the community.

Critical weaknesses

W1. I find limited novelty in the claimed contributions. Two of the four mentioned contributions are experimental evaluations of the method rather than methodological contributions. In my opinion, only the developed dataset, and potentially the proposed model, could be considered contributions. On the model side, the architecture seems particularly straightforward, considering that you propose embedding averaging of the GraphSAGE-based model. You should better contextualize and compare your model to other symmetry-aware methods, and clarify what the technical innovations are beyond applying known symmetry-enforcing methods, for example [1].

On the other hand, even though the dataset may be a useful asset for the wind engineering community, considering that there are only 66 geometries and 7 flow conditions, it is difficult to consider it a potential geometric deep learning benchmark for wind engineering. Besides, all the simulations involve single structures, while a much more interesting case might be the interactions between buildings in urban areas.

W2. The baselines used for comparison with WindMiL could be stronger. For example, the use of a simple MLP and a single GNN-style model is not enough to claim superiority. There are several stronger models in the literature, both mesh-based and point-cloud-based, that could possibly outperform those baselines. For example, [2, 3, 4, 5]. I think a comparison against these models would be better.

[1] Puny, Omri, et al. "Frame averaging for invariant and equivariant network design." arXiv preprint arXiv:2110.03336 (2021).

[2] Pfaff, Tobias, et al. "Learning mesh-based simulation with graph networks." arXiv preprint arXiv:2010.03409 (2020).

[3] Luo, Huakun, et al. "Transolver++: An accurate neural solver for pdes on million-scale geometries." arXiv preprint arXiv:2502.02414 (2025).

[4] Alkin, Benedikt, et al. "AB-UPT: Scaling neural CFD surrogates for high-fidelity automotive aerodynamics simulations via anchored-branched universal physics transformers." arXiv preprint arXiv:2502.09692 (2025).

[5] Adams, Corey, et al. "Geotransolver: Learning physics on irregular domains using multi-scale geometry aware physics attention transformer." arXiv preprint arXiv:2512.20399 (2025).

Questions for authors

Q1. What is the novelty of your proposed symmetry-based architecture compared with existing approaches?

Q2. Could you provide comparisons with stronger and more recent architectures, such as Transolver, GeoTransolver, AB-UPT, and Transolver++?

Q3. How do you expect the model to generalize to geometries beyond the 3 types of roof variation?

Q4. Could you please include relative error metrics such as L2 and L1 in addition to the current absolute metrics? I consider these more informative of the percentage error for each model with respect to the targets.

Q5. Your comparison with the GraphSAGE model using data augmentation seems unfair, as your model is called twice while GraphSAGE is called only once. I think a fairer comparison would be to average the predictions for both symmetric cases. This would help clarify whether the proposed latent averaging provides any advantage over a simpler symmetry-enforcement strategy. The same for any other baseline.

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

This paper provides a computational fluid dynamics benchmark based on large-eddy simulation (LES) for 462 low-rise buildings under atmospheric boundary layer flow, covering various roof shapes and wind directions.

Meanwhile, the authors propose WINDMIL, a reflection-equivariant GNN that learns to predict surface wind-load pressure. Specifically, the model takes point positions and outward unit normals as node features. For data augmentation, the authors add reflected positions and their corresponding outward unit normals to the node features used for prediction.

Critical strengths

The high-quality data provided in this paper is highly valuable. Each simulation under the LES constraints provides important data for studying these challenging problems.

The proposed model explicitly considers reflection symmetry, which is a meaningful and well-motivated design for this problem.

The provided benchmark offers a strong foundation for applying ML methods to these important real-world problems. It also creates valuable opportunities for further research, such as evaluating different regions of building surfaces separately to better understand the varying difficulty of pressure prediction.

Critical weaknesses

Discussion of symmetry. While I appreciate the high-quality dataset and the consideration of reflection symmetry, symmetry has also been widely studied in molecular modeling, such as in PaiNN and other equivariant GNNs. It would be valuable to briefly discuss the relationship between reflection symmetry and other symmetries, such as SO(3) and SE(3). However, I understand that the computational cost of these equivariant models may make them less suitable for dense point clouds in this task. A discussion of these trade-offs would further strengthen the paper.

Computational efficiency and sampling. Dense point clouds can introduce substantial computational costs during both training and inference, especially when considering MPNNs and symmetry constraints. Sampling-based surrogate methods are commonly considered to improve efficiency. Would the authors mind sharing an efficiency comparison or discussion of the proposed method with and without sampling? Additionally, if sampling is applied, would edges or regions with complex geometric features become more important for selecting sample points? Could the sampling strategy be adapted to the input building shapes? These would be interesting directions to explore and could further improve the practical applicability of the proposed method.

Questions for authors

Just out of curiosity, could the model be trained with additional features, such as wind direction and wind speed, to improve its generalizability to a broader range of settings? This could be an interesting direction for extending the applicability of the proposed method

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

Related papers

Discussion (0)

Sign in to comment.