pith. machine review for the scientific record. sign in

arxiv: 2603.21210 · v2 · submitted 2026-03-22 · 💻 cs.LG · cs.CE

Recognition: 2 theorem links

· Lean Theorem

Pretrained Video Models as Differentiable Physics Simulators for Urban Wind Flows

Janne Perini , Rafael Bischof , Moab Arar , Ay\c{c}a Duran , Michael A. Kraus , Siddhartha Mishra , Bernd Bickel
Authors on Pith no claims yet

Pith reviewed 2026-05-15 06:57 UTC · model grok-4.3

classification 💻 cs.LG cs.CE
keywords video diffusion modelsdifferentiable surrogatesurban wind flowsCFD accelerationinverse optimizationbuilding layout designwind comfortphysics-informed fine-tuning
0
0 comments X

The pith

A repurposed video diffusion model acts as a fast differentiable surrogate for urban wind flow simulations and enables direct gradient-based optimization of building positions.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper demonstrates that fine-tuning a large pretrained video model on simulated wind data around city buildings produces a surrogate capable of generating full time sequences of wind flows rapidly. This surrogate replaces costly traditional fluid dynamics calculations while remaining fully differentiable, allowing building layouts to be adjusted automatically through backpropagation to enhance wind safety and pedestrian comfort. Experiments confirm that the approach works for both single and multiple wind inlets, with resulting designs validated against accurate CFD runs. A sympathetic reader would care because it opens the door to exploring far more urban design options than current methods permit without sacrificing physical fidelity.

Core claim

Starting from the LTX-Video latent video transformer, fine-tuning on 10,000 procedurally generated 2D incompressible CFD cases with a physics-informed decoder loss yields WinDiNet, which produces 112-frame wind rollouts in under a second and supports end-to-end differentiable inverse optimization of urban footprints for improved wind metrics, with all discovered improvements confirmed by ground-truth CFD.

What carries the argument

WinDiNet, the fine-tuned latent video diffusion model that serves as an end-to-end differentiable surrogate for time-resolved 2D incompressible CFD wind simulations around building layouts.

If this is right

  • Full 112-frame wind flow predictions become available in under one second instead of hours of traditional computation.
  • Building positions can be adjusted via backpropagation to satisfy multiple wind comfort objectives simultaneously.
  • The surrogate outperforms purpose-built neural PDE solvers on the tested 2D incompressible cases.
  • Effective layouts are discovered for both single-inlet and multi-inlet urban configurations.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same fine-tuning strategy may transfer to other time-dependent physics problems that admit video-like representations.
  • If systematic biases remain small, the method could be extended to 3D flows or coupled with real sensor data for calibration.
  • Embedding the surrogate in interactive design software would allow rapid iteration over thousands of candidate urban layouts.

Load-bearing premise

Fine-tuning a general video model on procedurally generated 2D cases yields predictions accurate enough for gradient-based layout optimization without introducing biases that invalidate the discovered optima on real urban scenes.

What would settle it

An optimized layout produced by the surrogate that, when re-evaluated with full CFD, shows worse wind safety or comfort metrics than the starting layout.

Figures

Figures reproduced from arXiv: 2603.21210 by Ay\c{c}a Duran, Bernd Bickel, Janne Perini, Michael A. Kraus, Moab Arar, Rafael Bischof, Siddhartha Mishra.

Figure 1
Figure 1. Figure 1: Overview of the proposed framework. (a) Procedurally generated urban layouts are simulated with a 2D incompressible Euler solver to produce training data. (b) A latent diffusion model with a physics-informed VAE is trained to generate wind field sequences conditioned on building footprint, inlet speed uin, and domain size L. (c) At inference, the model generates horizontal and vertical velocity fields (u, … view at source ↗
Figure 2
Figure 2. Figure 2: Channel decomposition of a single simulation frame. From left to right: encoded RGB [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Representative samples from the training dataset. Each tile shows the RGB-encoded [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Wind speed magnitude predicted by Dec. FT Physics for a procedurally generated urban [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Learned channel transformation by the color adapter. VAE adaptation is performed as a separate stage before diffusion model training. To eval￾uate its effect in isolation, we encode ground￾truth velocity fields into the latent space and decode them back without involving the dif￾fusion model ( [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 5
Figure 5. Figure 5: VAE reconstruction quality at t=90 for a sample from the test set. all VAE variants, the gap between the Scalar Base and Scalar Dec. FT Physics (15.6% reduction in VRMSE) can be attributed to decoder adaptation. 4.5 Inference Configuration [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Grid search over CFG scale and number of denoising steps on the validation set (Scalar con￾ditioning, Dec. FT Physics). Cyan outline shows the best configuration. We select the number of denoising steps and the classifier-free guidance (CFG) scale via grid search on the validation set ( [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Inverse optimization pipeline, zoomed in for visibility. (a) Movable buildings (red) and [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: shows results for a single inlet boundary condition (left-to-right wind at 15 m/s) in rigid mode. The buildings translate to form a windbreak upstream of the objective region, deflecting the incoming flow and reducing through flow [PITH_FULL_IMAGE:figures/full_fig_p009_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Single-inlet morph optimization (15 m/s, left to right). Buildings are subdivided into independently movable sub-blocks that can deform the overall shape, targeting the 1–5 m/s band. Urban areas often experience wind from various directions throughout the year, and comfort re￾quirements may differ between seasons and climate zones. In winter, shelter from cold wind is desirable (low target speeds), wherea… view at source ↗
Figure 11
Figure 11. Figure 11: Multi-inlet rigid optimization with left inlet (15, 0) m/s (target comfort band 1–3 m/s) and top inlet (0, 15) m/s (target comfort band 3–5 m/s). A single layout is co-optimized for both inlets and evaluated under each direction separately. Acknowledgments and Disclosure of Funding We thank Cornelia Kalender for her guidance on selecting appropriate evaluation metrics for the CFD comparisons. We thank Kai… view at source ↗
Figure 12
Figure 12. Figure 12: shows the marginal speed distributions for both velocity components across the full training set. The horizontal component u is approximately uniform between 0 and 20 m/s, which is expected since the inlet flow is always aligned with the u-axis. The vertical component v is roughly normally distributed around zero, with most values confined to ±10 m/s. Both distributions exhibit a sharp peak at zero, likel… view at source ↗
Figure 13
Figure 13. Figure 13: Multi-inlet layout optimization: left (15, 0) m/s (comfort band 1–3 m/s) and top (0, 15) m/s (comfort band 3–5 m/s). Rigid mode (a–f) translates buildings, morph mode (g–l) additionally deforms them. E.9 Surrogate Comparison: WinDiNet vs. OFormer To assess how surrogate quality affects inverse optimization, we repeat the single-inlet rigid ex￾periment using OFormer—a competitive baseline in our forward-pr… view at source ↗
Figure 14
Figure 14. Figure 14: Inverse optimization using OFormer as surrogate (single inlet, rigid mode, Fig. 9). [PITH_FULL_IMAGE:figures/full_fig_p023_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Surrogate vs. ground-truth comfort loss over optimization steps for all five inverse design [PITH_FULL_IMAGE:figures/full_fig_p024_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Predicted velocity fields at frame 56 for four real-world urban configurations (left to right): [PITH_FULL_IMAGE:figures/full_fig_p025_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Temporal extrapolation beyond the T=112-frame training horizon. Rows show predictions at rollout lengths T, 2T, and 4T. The bottom row is ground-truth CFD. All predictions use a single forward pass with no autoregressive chaining [PITH_FULL_IMAGE:figures/full_fig_p026_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Per-timestep VRMSE for rollout lengths 1× (T=112), 2× (2T=224), and 4× (4T=448) the training horizon. Shaded regions indicate frames beyond each model’s generation length. 26 [PITH_FULL_IMAGE:figures/full_fig_p026_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: VRMSE and spectral divergence as a function of [PITH_FULL_IMAGE:figures/full_fig_p027_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Wind speed magnitude at t = 24 for inlet speeds from 5 to 29 m/s (left to right). Top row: ground truth; bottom row: model prediction. Out-of-distribution inlet speeds (uin > 20 m/s) are highlighted with a cyan border. F.4 Domain Size Generalization The training set covers L ∈ [900, 1400] m. We vary L from 700 to 1600 m in 100 m increments, with uin fixed at 18 m/s (the speed with lowest spectral divergen… view at source ↗
Figure 21
Figure 21. Figure 21: VRMSE and spectral divergence as a function of [PITH_FULL_IMAGE:figures/full_fig_p028_21.png] view at source ↗
read the original abstract

Designing urban spaces that provide pedestrian wind comfort and safety requires time-resolved Computational Fluid Dynamics (CFD) simulations, but their current computational cost makes extensive design exploration impractical. We introduce WinDiNet (Wind Diffusion Network), a pretrained video diffusion model that is repurposed as a fast, differentiable surrogate for this task. Starting from LTX-Video, a 2B-parameter latent video transformer, we fine-tune on 10,000 2D incompressible CFD simulations over procedurally generated building layouts. A systematic study of training regimes, conditioning mechanisms, and VAE adaptation strategies, including a physics-informed decoder loss, identifies a configuration that outperforms purpose-built neural PDE solvers. The resulting model generates full 112-frame rollouts in under a second. As the surrogate is end-to-end differentiable, it doubles as a physics simulator for gradient-based inverse optimization: given an urban footprint layout, we optimize building positions directly through backpropagation to improve wind safety as well as pedestrian wind comfort. Experiments on single- and multi-inlet layouts show that the optimizer discovers effective layouts even under challenging multi-objective configurations, with all improvements confirmed by ground-truth CFD simulations.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper introduces WinDiNet, a fine-tuned 2B-parameter latent video diffusion model (based on LTX-Video) repurposed as a differentiable surrogate for 2D incompressible CFD simulations of urban wind flows. Fine-tuned on 10,000 procedurally generated building layouts, the model produces full 112-frame rollouts in under one second and is applied to gradient-based inverse optimization of building positions to improve wind safety and pedestrian comfort, with discovered optima validated by ground-truth CFD.

Significance. If the surrogate's accuracy and gradient fidelity hold on real urban layouts, the work would provide a practical route to rapid, end-to-end differentiable physics simulation for urban design, enabling extensive inverse optimization that remains computationally prohibitive with conventional CFD. The demonstration that a pretrained video model can be adapted into a physics surrogate with physics-informed losses is a notable contribution to scientific machine learning.

major comments (3)
  1. [Abstract / Experiments] Abstract and Experiments section: the claim that the model 'outperforms purpose-built neural PDE solvers' is unsupported by any reported quantitative error metrics (e.g., velocity-field RMSE, divergence error, or rollout accuracy over 112 frames) or ablation tables comparing training regimes, conditioning mechanisms, and the physics-informed decoder loss. Without these numbers the superiority assertion and the suitability for gradient-based optimization cannot be evaluated.
  2. [Methods / Data generation] Training data and generalization discussion: the 10,000 procedurally generated 2D layouts are described only at a high level; no statistics on shape diversity, aspect-ratio distribution, inlet variability, or boundary-layer fidelity are supplied. This leaves open the possibility that the learned dynamics contain systematic biases that would distort the loss landscape during building-position optimization, exactly as flagged by the skeptic note.
  3. [Optimization results] Optimization experiments: while the abstract states that 'all improvements [are] confirmed by ground-truth CFD,' neither the number of optimized layouts, the magnitude of the reported wind-speed or comfort gains, nor any comparison against baseline layouts or random search is quantified. Post-hoc verification on a small undisclosed set does not establish that the discovered optima are robust to surrogate artifacts.
minor comments (2)
  1. [Introduction] The acronym WinDiNet is introduced without an explicit expansion or architectural diagram in the early sections; a single figure showing the conditioning and VAE adaptation pipeline would improve readability.
  2. [Results] Clarify whether the 112-frame rollout length is fixed by the video model architecture or chosen for the CFD task, and report wall-clock times on the specific hardware used for both inference and back-propagation.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their insightful and constructive comments on our manuscript. We are pleased that the referee recognizes the potential of WinDiNet as a practical approach for differentiable physics simulation in urban design. We address each of the major comments below and will incorporate the suggested revisions to strengthen the paper.

read point-by-point responses

  1. Referee: [Abstract / Experiments] Abstract and Experiments section: the claim that the model 'outperforms purpose-built neural PDE solvers' is unsupported by any reported quantitative error metrics (e.g., velocity-field RMSE, divergence error, or rollout accuracy over 112 frames) or ablation tables comparing training regimes, conditioning mechanisms, and the physics-informed decoder loss. Without these numbers the superiority assertion and the suitability for gradient-based optimization cannot be evaluated.

    Authors: We fully agree that explicit quantitative metrics are necessary to support the claim of outperforming purpose-built neural PDE solvers and to confirm suitability for gradient-based optimization. In the revised manuscript, we will add detailed quantitative comparisons in the Experiments section, including velocity-field RMSE, divergence error, and rollout accuracy over 112 frames. Additionally, we will include ablation tables that compare different training regimes, conditioning mechanisms, and the impact of the physics-informed decoder loss. These metrics will demonstrate the performance gains and provide the necessary evidence for the assertions made. revision: yes

  2. Referee: [Methods / Data generation] Training data and generalization discussion: the 10,000 procedurally generated 2D layouts are described only at a high level; no statistics on shape diversity, aspect-ratio distribution, inlet variability, or boundary-layer fidelity are supplied. This leaves open the possibility that the learned dynamics contain systematic biases that would distort the loss landscape during building-position optimization, exactly as flagged by the skeptic note.

    Authors: We acknowledge the importance of providing more detailed dataset statistics to allow assessment of potential biases. In the revised Methods section, we will include comprehensive statistics on the procedural generation process, such as distributions of building shapes and sizes, aspect ratios, inlet flow variability, and boundary layer characteristics. This additional information will help evaluate the diversity of the training data and the robustness of the learned surrogate for optimization purposes. revision: yes

  3. Referee: [Optimization results] Optimization experiments: while the abstract states that 'all improvements [are] confirmed by ground-truth CFD,' neither the number of optimized layouts, the magnitude of the reported wind-speed or comfort gains, nor any comparison against baseline layouts or random search is quantified. Post-hoc verification on a small undisclosed set does not establish that the discovered optima are robust to surrogate artifacts.

    Authors: We agree that more quantitative details on the optimization experiments are required to demonstrate the robustness of the results. In the revised manuscript, we will specify the number of layouts optimized, quantify the improvements in wind-speed and comfort metrics, and provide comparisons to baseline configurations and random search. We will also clarify the extent of the ground-truth CFD validation, ensuring that the verification covers a representative set of cases to confirm that the optima are not artifacts of the surrogate. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation relies on external CFD data and independent verification

full rationale

The paper's core chain—pretraining on LTX-Video, fine-tuning on 10,000 external procedurally generated 2D CFD cases, producing differentiable rollouts, and performing gradient-based optimization with post-hoc ground-truth CFD confirmation—does not reduce any prediction or claim to its inputs by construction. No self-definitional loops, fitted parameters renamed as predictions, or load-bearing self-citations appear in the provided text. The optimization results are explicitly validated outside the surrogate, keeping the derivation self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

The central claim rests on the assumption that a video diffusion architecture pretrained on natural video can be adapted to incompressible fluid dynamics via fine-tuning on synthetic 2D CFD data, with the only added physics signal coming from a decoder loss term. No new physical constants or entities are introduced beyond the model weights.

free parameters (1)
  • fine-tuning hyperparameters and conditioning mechanisms
    The systematic study selects among training regimes, conditioning mechanisms, and VAE adaptation strategies; these choices are fitted to the 10,000 CFD simulations.
axioms (1)
  • domain assumption A 2D incompressible CFD simulation on procedurally generated building layouts is a sufficient proxy for real 3D urban wind flows.
    Invoked when claiming the surrogate is useful for urban design.
invented entities (1)
  • WinDiNet no independent evidence
    purpose: Name for the fine-tuned LTX-Video model used as wind surrogate
    New label for the repurposed model; no independent physical existence claimed.

pith-pipeline@v0.9.0 · 5523 in / 1579 out tokens · 43445 ms · 2026-05-15T06:57:29.953804+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

59 extracted references · 59 canonical work pages · 8 internal anchors

  1. [1]

    Stable Video Diffusion: Scaling Latent Video Diffusion Models to Large Datasets

    Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Do- minik Lorenz, Yam Levi, Zion English, Vikram V oleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets.arXiv preprint arXiv:2311.15127, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  2. [2]

    Eurocode 1: Actions on structures – part 1-4: General actions – wind actions

    CEN. Eurocode 1: Actions on structures – part 1-4: General actions – wind actions. Standard, 2010

    work page 2010

  3. [3]

    Generalization of urban wind environ- ment using fourier neural operator across different wind directions and cities.arXiv preprint arXiv:2501.05499, 2025

    Cheng Chen, Geng Tian, Shaoxiang Qin, Senwen Yang, Dingyang Geng, Dongxue Zhan, Jinqiu Yang, David Vidal, and Liangzhu Leon Wang. Generalization of urban wind environ- ment using fourier neural operator across different wind directions and cities.arXiv preprint arXiv:2501.05499, 2025

    work page arXiv 2025

  4. [4]

    Wind microclimate guidelines for developments in the city of london

    City of London Corporation. Wind microclimate guidelines for developments in the city of london. https://www.cityoflondon.gov.uk/assets/Services-Environment/ wind-microclimate-guidelines.pdf, 2019

    work page 2019

  5. [5]

    Deep learning for urban wind prediction: An mlp-mixer approach with 3d encoding.Building and Environment, page 113495, 2025

    Adam Clarke, Knut Erik Teigen Giljarhus, Luca Oggiano, Alistair Saddington, and Karthik Depuru-Mohan. Deep learning for urban wind prediction: An mlp-mixer approach with 3d encoding.Building and Environment, page 113495, 2025

    work page 2025

  6. [6]

    Alfredo Vicente Clemente, Knut Erik Teigen Giljarhus, Luca Oggiano, and Massimiliano Ruocco. Rapid pedestrian-level wind field prediction for early-stage design using pareto- optimized convolutional neural networks.Computer-Aided Civil and Infrastructure Engineering, 39(18):2826–2839, 2024

    work page 2024

  7. [7]

    Conditional neural field latent diffusion model for generating spatiotemporal turbulence.Nature Communi- cations, 15(1):10416, 2024

    Pan Du, Meet Hemant Parikh, Xiantao Fan, Xin-Yang Liu, and Jian-Xun Wang. Conditional neural field latent diffusion model for generating spatiotemporal turbulence.Nature Communi- cations, 15(1):10416, 2024

    work page 2024

  8. [8]

    Spencer Folk, John Melton, Benjamin W. L. Margolis, Mark Yim, and Vijay Kumar. Learning local urban wind flow fields from range sensing.IEEE Robotics and Automation Letters, 9(9): 7413–7420, 2024. doi: 10.1109/LRA.2024.3426209

    work page doi:10.1109/lra.2024.3426209 2024

  9. [9]

    Force prompting: Video generation models can learn and generalize physics-based control signals.arXiv preprint arXiv:2505.19386, 2025

    Nate Gillman, Charles Herrmann, Michael Freeman, Daksh Aggarwal, Evan Luo, Deqing Sun, and Chen Sun. Force prompting: Video generation models can learn and generalize physics-based control signals.arXiv preprint arXiv:2505.19386, 2025

    work page arXiv 2025

  10. [10]

    Generative urban flow modeling: From geometry to airflow with graph diffusion

    Francisco Giral, Álvaro Manzano, Ignacio Gómez, Petros Koumoutsakos, and Soledad Le Clainche. Generative urban flow modeling: From geometry to airflow with graph diffusion. arXiv preprint arXiv:2512.14725, 2025

    work page arXiv 2025

  11. [11]

    Efficient token mixing for transformers via adaptive fourier neural operators

    John Guibas, Morteza Mardani, Zongyi Li, Andrew Tao, Anima Anandkumar, and Bryan Catanzaro. Efficient token mixing for transformers via adaptive fourier neural operators. In International conference on learning representations, 2021

    work page 2021

  12. [12]

    LTX-Video: Realtime Video Latent Diffusion

    Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richard- son, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, et al. Ltx-video: Realtime video latent diffusion.arXiv preprint arXiv:2501.00103, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  13. [13]

    Poseidon: Efficient foundation models for pdes

    Maximilian Herde, Bogdan Raoni ´c, Tobias Rohner, Roger Käppeli, Roberto Molinaro, Em- manuel De Bezenac, and Siddhartha Mishra. Poseidon: Efficient foundation models for pdes. Advances in Neural Information Processing Systems, 37:72525–72624, 2024

    work page 2024

  14. [14]

    Video diffusion models.Advances in neural information processing systems, 35: 8633–8646, 2022

    Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models.Advances in neural information processing systems, 35: 8633–8646, 2022

    work page 2022

  15. [15]

    Lora: Low-rank adaptation of large language models.Iclr, 1(2):3, 2022

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Liang Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models.Iclr, 1(2):3, 2022. 12

    work page 2022

  16. [16]

    Accelerated environmental performance-driven urban design with generative adversarial network.Building and Environment, 224:109575, 2022

    Chenyu Huang, Gengjia Zhang, Jiawei Yao, Xiaoxin Wang, John Kaiser Calautit, Cairong Zhao, Na An, and Xi Peng. Accelerated environmental performance-driven urban design with generative adversarial network.Building and Environment, 224:109575, 2022

    work page 2022

  17. [17]

    The ground level wind environment in built-up areas

    N Isyumov. The ground level wind environment in built-up areas. Inproc. of 4ˆ< th> Int. Conf. on Wind Effects on Buildings and Structures, London, 1975, 1975

    work page 1975

  18. [18]

    Towards cfd-based optimization of urban wind condi- tions: Comparison of genetic algorithm, particle swarm optimization, and a hybrid algorithm

    Zeynab Kaseb and Morteza Rahbar. Towards cfd-based optimization of urban wind condi- tions: Comparison of genetic algorithm, particle swarm optimization, and a hybrid algorithm. Sustainable Cities and Society, 77:103565, 2022

    work page 2022

  19. [19]

    A gan-based surrogate model for instantaneous urban wind flow prediction.Building and Environment, 242:110384, 2023

    Patrick Kastner and Timur Dogan. A gan-based surrogate model for instantaneous urban wind flow prediction.Building and Environment, 242:110384, 2023

    work page 2023

  20. [20]

    Adam: A Method for Stochastic Optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980, 2014

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  21. [21]

    T. V . Lawson. The wind content of the built environment.Journal of Wind Engineering and Industrial Aerodynamics, 3(2–3):93–105, 1978

    work page 1978

  22. [22]

    High resolution large-eddy simulation of turbulent flow around buildings, 2007

    Marcus Oliver Letzel. High resolution large-eddy simulation of turbulent flow around buildings, 2007

    work page 2007

  23. [23]

    Videopde: Unified generative pde solving via video inpainting diffusion models.arXiv preprint arXiv:2506.13754, 2025

    Edward Li, Zichen Wang, Jiahe Huang, and Jeong Joon Park. Videopde: Unified generative pde solving via video inpainting diffusion models.arXiv preprint arXiv:2506.13754, 2025

    work page arXiv 2025

  24. [24]

    Transformer for partial differential equations’ operator learning.arXiv preprint arXiv:2205.13671, 2022

    Zijie Li, Kazem Meidani, and Amir Barati Farimani. Transformer for partial differential equations’ operator learning.arXiv preprint arXiv:2205.13671, 2022

    work page arXiv 2022

  25. [25]

    Fourier Neural Operator for Parametric Partial Differential Equations

    Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differen- tial equations.arXiv preprint arXiv:2010.08895, 2020

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  26. [26]

    Soft rasterizer: A differentiable renderer for image-based 3d reasoning

    Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft rasterizer: A differentiable renderer for image-based 3d reasoning. InProceedings of the IEEE/CVF international conference on computer vision, pages 7708–7717, 2019

    work page 2019

  27. [27]

    Accurate and efficient urban wind prediction at city-scale with memory-scalable graph neural network.Sustainable Cities and Society, 2023

    Zhijian Liu, Siqi Zhang, Xuqiang Shao, and Zhaohui Wu. Accurate and efficient urban wind prediction at city-scale with memory-scalable graph neural network.Sustainable Cities and Society, 2023

    work page 2023

  28. [28]

    Tipping point forecasting in non-stationary dynamics on function spaces.arXiv preprint arXiv:2308.08794, 2023

    Miguel Liu-Schiaffini, Clare E Singer, Nikola Kovachki, Tapio Schneider, Kamyar Azizzade- nesheli, and Anima Anandkumar. Tipping point forecasting in non-stationary dynamics on function spaces.arXiv preprint arXiv:2308.08794, 2023

    work page arXiv 2023

  29. [29]

    Decoupled Weight Decay Regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  30. [30]

    Machine learning predicts pedestrian wind flow from urban morphology and prevailing wind direction

    Jiachen Lu, Wei Li, Sanaa Hobeichi, Shakir Aymam Azad, and Negin Nazarian. Machine learning predicts pedestrian wind flow from urban morphology and prevailing wind direction. Environmental Research Letters, 20(5):054006, 2025

    work page 2025

  31. [31]

    Cfd modeling of micro and urban climates: Problems to be solved in the new decade.Sustainable Cities and Society, 69:102839, 2021

    Parham A Mirzaei. Cfd modeling of micro and urban climates: Problems to be solved in the new decade.Sustainable Cities and Society, 69:102839, 2021

    work page 2021

  32. [32]

    Pedestrian wind factor estimation in complex urban environments

    Sarah Mokhtar, Matt Beveridge, Yumeng Cao, and Iddo Drori. Pedestrian wind factor estimation in complex urban environments. InAsian conference on machine learning, pages 486–501. PMLR, 2021

    work page 2021

  33. [33]

    Generative ai for fast and accurate statistical computation of fluids.arXiv preprint arXiv:2409.18359, 2024

    Roberto Molinaro, Samuel Lanthaler, Bogdan Raoni ´c, Tobias Rohner, Victor Armegioiu, Stephan Simonis, Dana Grund, Yannick Ramic, Zhong Yi Wan, Fei Sha, et al. Generative ai for fast and accurate statistical computation of fluids.arXiv preprint arXiv:2409.18359, 2024

    work page arXiv 2024

  34. [34]

    Physix: A foundation model for physics simulations.arXiv preprint arXiv:2506.17774, 2025

    Tung Nguyen, Arsh Koneru, Shufan Li, and Aditya Grover. Physix: A foundation model for physics simulations.arXiv preprint arXiv:2506.17774, 2025. 13

    work page arXiv 2025

  35. [35]

    PhysVideo- Generator: Towards Physically Aware Video Generation via Latent Physics Guidance.arXiv e-prints, art

    Siddarth Nilol Kundur Satish, Devesh Jaiswal, Hongyu Chen, and Abhishek Bakshi. PhysVideo- Generator: Towards Physically Aware Video Generation via Latent Physics Guidance.arXiv e-prints, art. arXiv:2601.03665, January 2026. doi: 10.48550/arXiv.2601.03665

    work page doi:10.48550/arxiv.2601.03665 2026

  36. [36]

    NVIDIA PhysicsNeMo: An open-source framework for physics-ML model building and training, 2025

    NVIDIA Corporation. NVIDIA PhysicsNeMo: An open-source framework for physics-ML model building and training, 2025. URLhttps://github.com/NVIDIA/physicsnemo

    work page 2025

  37. [37]

    The well: a large-scale collection of diverse physics simulations for machine learning.Advances in Neural Information Processing Systems, 37:44989–45037, 2024

    Ruben Ohana, Michael McCabe, Lucas Meyer, Rudy Morel, Fruzsina J Agocs, Miguel Beneitez, Marsha Berger, Blakesley Burkhart, Stuart B Dalziel, Drummond B Fielding, et al. The well: a large-scale collection of diverse physics simulations for machine learning.Advances in Neural Information Processing Systems, 37:44989–45037, 2024

    work page 2024

  38. [38]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. InProceedings of the IEEE/CVF international conference on computer vision, pages 4195–4205, 2023

    work page 2023

  39. [39]

    Film: Visual reasoning with a general conditioning layer

    Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. InProceedings of the AAAI conference on artificial intelligence, volume 32, 2018

    work page 2018

  40. [40]

    Modeling multivariable high-resolution 3d urban microclimate using localized fourier neural operator.Building and Environment, 273:112668, 2025

    Shaoxiang Qin, Dongxue Zhan, Dingyang Geng, Wenhui Peng, Geng Tian, Yurong Shi, Naiping Gao, Xue Liu, and Liangzhu Leon Wang. Modeling multivariable high-resolution 3d urban microclimate using localized fourier neural operator.Building and Environment, 273:112668, 2025

    work page 2025

  41. [41]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. InInternational Conference on Medical image computing and computer-assisted intervention, pages 234–241. Springer, 2015

    work page 2015

  42. [42]

    En-Ze Rui, Zheng-Wei Chen, Yi-Qing Ni, Lei Yuan, and Guang-Zhi Zeng. Reconstruction of 3d flow field around a building model in wind tunnel: a novel physics-informed neural network framework adopting dynamic prioritization self-adaptive loss balance strategy.Engineering Applications of Computational Fluid Mechanics, 17(1):2238849, 2023

    work page 2023

  43. [43]

    Temporal generative adversarial nets with singular value clipping

    Masaki Saito, Eiichi Matsumoto, and Shunta Saito. Temporal generative adversarial nets with singular value clipping. InProceedings of the IEEE international conference on computer vision, pages 2830–2839, 2017

    work page 2017

  44. [44]

    Pignn-cfd: A physics- informed graph neural network for rapid predicting urban wind field defined on unstructured mesh.Building and Environment, 232:110056, 2023

    Xuqiang Shao, Zhijian Liu, Siqi Zhang, Zijia Zhao, and Chenxing Hu. Pignn-cfd: A physics- informed graph neural network for rapid predicting urban wind field defined on unstructured mesh.Building and Environment, 232:110056, 2023

    work page 2023

  45. [45]

    Pedestrian wind comfort analysis – online documentation

    SimScale GmbH. Pedestrian wind comfort analysis – online documentation. https://www. simscale.com/docs/analysis-types/pedestrian-wind-comfort-analysis/, 2025

    work page 2025

  46. [46]

    A hierarchical deep learning model for predicting pedestrian-level urban winds.Building and Environment, page 114354, 2026

    Reda Snaiki, Jiachen Lu, Shaopeng Li, and Negin Nazarian. A hierarchical deep learning model for predicting pedestrian-level urban winds.Building and Environment, page 114354, 2026

    work page 2026

  47. [47]

    Towards a foundation model for partial differential equations across physics domains

    Eduardo Soares, Emilio Vital Brazil, Victor Shirasuna, Breno WSR de Carvalho, and Cristiano Malossi. Towards a foundation model for partial differential equations across physics domains. arXiv preprint arXiv:2511.21861, 2025

    work page arXiv 2025

  48. [48]

    Pdebench: An extensive benchmark for scientific machine learning.Advances in neural information processing systems, 35:1596–1611, 2022

    Makoto Takamoto, Timothy Praditia, Raphael Leiteritz, Daniel MacKinlay, Francesco Alesiani, Dirk Pflüger, and Mathias Niepert. Pdebench: An extensive benchmark for scientific machine learning.Advances in neural information processing systems, 35:1596–1611, 2022

    work page 2022

  49. [49]

    Fourier features let networks learn high frequency functions in low dimensional domains.Advances in neural information processing systems, 33:7537–7547, 2020

    Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains.Advances in neural information processing systems, 33:7537–7547, 2020

    work page 2020

  50. [50]

    MAGI-1: Autoregressive Video Generation at Scale

    Hansi Teng, Hongyu Jia, Lei Sun, Lingzhi Li, Maolin Li, Mingqiu Tang, Shuai Han, Tianning Zhang, WQ Zhang, Weifeng Luo, et al. Magi-1: Autoregressive video generation at scale.arXiv preprint arXiv:2505.13211, 2025. 14

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  51. [51]

    Mocogan: Decomposing motion and content for video generation

    Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 1526–1535, 2018

    work page 2018

  52. [52]

    Diffusion models are real-time game engines.arXiv preprint arXiv:2408.14837, 2024

    Dani Valevski, Yaniv Leviathan, Moab Arar, and Shlomi Fruchter. Diffusion models are real-time game engines.arXiv preprint arXiv:2408.14837, 2024

    work page arXiv 2024

  53. [53]

    Generating videos with scene dynamics

    Carl V ondrick, Hamed Pirsiavash, and Antonio Torralba. Generating videos with scene dynamics. Advances in neural information processing systems, 29, 2016

    work page 2016

  54. [54]

    Wan: Open and Advanced Large-Scale Video Generative Models

    Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, et al. Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  55. [55]

    Physctrl: Generative physics for controllable and physics-grounded video generation.arXiv preprint arXiv:2509.20358, 2025

    Chen Wang, Chuhao Chen, Yiming Huang, Zhiyang Dou, Yuan Liu, Jiatao Gu, and Lingjie Liu. Physctrl: Generative physics for controllable and physics-grounded video generation.arXiv preprint arXiv:2509.20358, 2025

    work page arXiv 2025

  56. [56]

    Video models are zero-shot learners and reasoners

    Thaddäus Wiedemer, Yuxuan Li, Paul Vicol, Shixiang Shane Gu, Nick Matarese, Kevin Swersky, Been Kim, Priyank Jaini, and Robert Geirhos. Video models are zero-shot learners and reasoners. arXiv preprint arXiv:2509.20328, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  57. [57]

    InThe Thirteenth International Conference on Learning Representations(2025)

    Florian Wiesner, Matthias Wessling, and Stephen Baek. Towards a physics foundation model. arXiv preprint arXiv:2509.13805, 2025

    work page arXiv 2025

  58. [58]

    A review of surrogate-assisted design optimization for improving urban wind environment.Building and Environment, 253:111157, 2024

    Yihan Wu and Steven Jige Quan. A review of surrogate-assisted design optimization for improving urban wind environment.Building and Environment, 253:111157, 2024

    work page 2024

  59. [59]

    Think before you diffuse: Infusing physical rules into video diffusion.arXiv preprint arXiv:2505.21653, 2025

    Ke Zhang, Cihan Xiao, Jiacong Xu, Yiqun Mei, and Vishal M Patel. Think before you diffuse: Infusing physical rules into video diffusion.arXiv preprint arXiv:2505.21653, 2025. 15 A Dataset Details A.1 Incompressible CFD Setup We model the atmospheric boundary-layer flow as an incompressible fluid, which is appropriate for typical urban wind speeds where Ma...

    work page arXiv 2025