PaperlayerView exact revision on arXiv

Source-linked AI summary

Conditional Flow Matching for Generation of 3D Multi-variable Instantaneous Urban Microclimate Fields

Peng Liu, Shaoxiang Qin, Theodore Potsis, Lili Ji, Dingyang Geng, Liangzhu Leon Wang

arXiv:2610.10430v1physics.flu-dyncs.CEcs.LGphysics.ao-ph

TL;DR

Urban microclimate applications need instantaneous turbulent fields, but LES is costly and deterministic surrogates cannot represent stochastic variability. The paper uses CFM conditioned on building geometry and mean flow to generate 3D instantaneous fields. It reports an efficient route to probabilistic assessment and gust estimation at a fraction of full LES cost.

  • Problem

    Urban applications need stochastic instantaneous wind information, while LES is computationally costly and deterministic regressors tend to average out turbulent fluctuations.

    Source: §1 · p. 2

  • Method

    CFM generates 3D instantaneous velocity and temperature fields conditioned on building geometry and steady-state mean flow, using overlapping patches with shared-noise initialization for full-domain reconstruction.

    Source: §1 · p. 4, §5 · p. 35

  • Results

    CFM provides an efficient path to stochastic urban microclimate fields; one snapshot takes seconds versus hours for CityFFD, and 35 take under 10 minutes versus hours.

    Source: §4 · p. 31, §5 · p. 35

  • Takeaways & Limitations

    The generated ensembles support probabilistic wind and temperature comfort assessment and gust estimation at a fraction of the cost of full LES campaigns.

    Source: §5 · p. 35

    In dense street canyons, CFM underestimates ensemble variance and vertical-velocity fluctuations, with NRMSE 26.56% and r2 0.029.

    Source: §3.6 · p. 30

Abstract

from arXiv · show

Rapid and accurate prediction of urban wind and temperature fields is important for urban microclimate design and climate adaptation. Large-eddy simulation (LES) effectively resolves these instantaneous fields, but its application is limited in iterative design of urban microclimate applications due to high computational cost. Existing regressive data-driven models offers quick outputs, but they produce only deterministic point predictions that inherently fail to represent turbulent stochasticity. This paper adopts a novel generative framework of Conditional Flow Matching (CFM) that uses building geometry and mean flow as guidance to generate plausible three-dimensional instantaneous velocity and temperature fields for urban microclimate in seconds. To overcome the GPU memory bottleneck of pixel space 3D generation, the model operates in parallel on overlapping pixel space through a shared-noise initialization that preserves high spatial continuity of flow structure across the entire domain. Against reference LES data, the CFM surrogate can rapidly and accurately restore the first-order statistics with Normalized Root Mean Square Error (NRMSE) of 2.99% for wind and 1.77% for temperature, second-order turbulence metrics with NRMSE of 7.17% for wind and 8.84% for temperature, turbulent kinetic energy with NRMSE of 7%, probability density function and vertical profiles in representative locations. Wind engineering application of local gust prediction demonstrate that the speed and accuracy of CFM, supporting the use of generative AI for making turbulence-aware resilient urban design and climate adaptation more computationally feasible.

1. Introduction

The paper proposes Conditional Flow Matching (CFM) to generate instantaneous urban wind and temperature fields, addressing the cost of LES and the inability of deterministic regressors to represent turbulent variability. Its introduction frames scalable, spatially faithful 3D generation as a remaining challenge.

1. Introduction

  • Urban growth alters aerodynamics and heat transfer, while urban microclimate affects pedestrian comfort, ventilation, pollutant dispersion, and safety.The introduction highlights the Urban Heat Island effect and the need to understand instantaneous turbulent structures.Source: §1 · p. 2
  • LES resolves stochastic flow fluctuations but can require hours to days of CPU time for one large urban scenario, with probabilistic analysis multiplying the cost.Deterministic regression models are fast but tend to average out stochastic fluctuations and do not provide full instantaneous flow states.Source: §1 · p. 2
  • Pixel-space generation preserves local flow detail but is constrained by GPU resources, whereas latent-space compression can blur high-frequency structures.The paper identifies scalable generation of large transient 3D urban fields with spatial fidelity and global consistency as an unresolved gap.Source: §1 · pp. 3–4
  • CFM generates instantaneous 3D velocity and temperature snapshots conditioned on building geometry and steady-state mean flow.The paper presents this as an application of CFM to urban microclimate.Source: §1 · p. 4
  • The framework aims to overcome GPU-memory limits for probabilistic generation in large 3D tensors with multiple variables.The introduction presents this as a contribution toward scalable methods.Source: §1 · p. 4
  • The paper targets this gap with conditional generation guided by geometry and mean flow, enabling plausible transient ensembles beyond mean-flow prediction alone.The stated applications include extreme-case analysis and real-world problems.Source: §1 · p. 4

2. Methodology

The methodology builds a CFM surrogate from CityFFD simulations, conditions instantaneous-field generation on urban geometry and mean flow, and reconstructs full-domain outputs from overlapping patches. Its evaluation compares generated ensembles with CityFFD across statistical measures and representative locations.

2.1 Numerical simulations

  • Twenty diverse urban morphologies provide 500 instantaneous snapshots each; simulations are cropped to 1.2 km × 1.2 km × 240 m subdomains and split into 17 training and 3 test cities.The cropped tensors are resampled to 8 m × 8 m × 3 m; snapshots are saved at 5-second intervals.Source: §2.1 · pp. 5, 6, 8

2.2 CFM framework

  • CFM learns an ODE velocity field along a linear optimal-transport path, conditioning transient-field generation on building geometry and time-averaged flow.Inputs include a building mask, signed distance field, mean-flow channels, and roof-height map; a 3D U-Net predicts four velocity and temperature channels.Source: §2.2 · pp. 8–9

2.3 Patch-based 3D generation and domain reconstruction

  • Overlapping 64³ patches address full-volume GPU memory limits; shared-noise initialization and weighted merging at each ODE step maintain coherence across patch boundaries.The inference loop transports patches with RK4 and merges overlaps using distance-based weights and a raised 3D cosine kernel.Source: §2.3 · pp. 11–13, Figure 4 · p. 12

2.4 Evaluation metrics

  • Evaluation checks ensemble convergence, first- and second-order statistics, turbulent kinetic energy, probability densities, and vertical profiles against CityFFD references.Vertical profiles are compared at an upstream site, a street canyon, and an open area; the framework also uses localized bias and correlation measures.Source: §2.4 · p. 13, b. Flow field statistics · pp. 14–15, c. Error metrics · p. 16

3. Results

Across the test urban configuration, the 35-snapshot CFM ensemble reproduces mean flow and temperature, turbulent statistics, PDFs, and representative vertical profiles, while local errors remain in highly turbulent regions.

3.1 Convergence of generated ensemble

  • Mean change stabilizes below 0.01% at N = 35, whereas standard-deviation change reaches that threshold at N = 70; the study uses 35 snapshots.The ensemble size balances numerical accuracy and inference speed.Source: Figure 5 · p. 17, §3.1 · p. 17

3.2 Instantaneous flow generations

  • Generated fields retain geometry-driven channels, stagnation zones, wakes, and thermal patterns while varying across plausible instantaneous realizations.Snapshots take 10–13 seconds each; overlapping 64^3 patches with shared-noise initialization and weighted merging yield spatially continuous fields.Source: §3.2 · pp. 18–20, Figure 6 · p. 19

3.3 Mean microclimate generation

  • The ensemble mean matches CityFFD, with velocity-component NRMSE around 1.925% and temperature NRMSE 1.773%.Mean velocity structures and thermal plumes are reproduced; the largest discrepancies are localized near building interfaces.Source: §3.3 · pp. 21, 23, Figure 9 · p. 22, Table 1 · p. 23

3.4 Turbulence generation

  • Standard-deviation NRMSE is 7.17% for u and 8.84% for temperature; TKE range-normalized RMSE is 7.052%, with its maximum below CityFFD’s.The CFM TKE maximum is 4.608 m²/s² versus CityFFD’s 5.049 m²/s²; standard-deviation errors cluster around building edges and shear layers.Source: §3.4 · pp. 24, 26, 27, Figure 10 · p. 25, Table 3 · p. 26

3.5 Probability density function

  • CFM PDFs largely overlap CityFFD distributions across four variables and four heights, but near-ground w overestimates peak density and misses some multi-peak structure.The near-surface discrepancy is associated with localized terrain-induced turbulence.Source: §3.5 · p. 27, Figure 12 · p. 28

3.6 Vertical wind profiles

  • Vertical profiles capture u and temperature across sites, including upstream u with r² = 0.996 and NRMSE 2.21%, but canyon w has NRMSE 26.56% and r² = 0.029.The authors caution against relying on localized vertical-velocity variance for fatigue design or peak wind loads inside deep street canyons.Source: §3.6 · pp. 28, 30, Figure 13 · p. 29, Table 4 · p. 30

4. Discussion

The discussion presents CFM ensembles as a route to probabilistic urban-wind assessment and reports fast generation with useful agreement to reference gust simulations. It also notes some peak-fluctuation smoothing and identifies extensions needed for continuous time series and broader urban geometries.

4. Discussion

  • CFM’s 35-snapshot ensemble supports probabilistic assessments such as pedestrian gust-exceedance maps and ventilation-rate estimates with uncertainty bounds.The discussion contrasts these analyses with deterministic surrogates’ single predictions.Source: §4 · p. 30
  • The model can act as a low-pass filter, occasionally underestimating peak fluctuations in high-fluctuation zones, although PDF tails remain well preserved.Standard-deviation differences are 0.35 m/s and 0.33 °C; the authors flag peak-gust applications as a concern.Source: §4 · p. 31
  • Shared-noise patch generation preserves spatial continuity and turbulent energy across boundaries; one snapshot takes seconds and 35 require under 10 minutes.This is compared with hours for CityFFD.Source: §4 · p. 31
  • Gust comparisons use a quasi-steady peak estimate based on mean and standard deviation; the study focuses on environmental wind, not wind-induced pressures.The approximate Gaussian assumption gives g = 3 as a three-sigma upper estimate.Source: §4 · p. 31
  • In vertical gust planes, CFM captures tallest-building wake recovery with differences mainly below 0.5 m/s, while canyon gust errors remain mainly below 0.6 m/s.The tallest-building differences are larger downstream and at lower heights; canyon gusts show a wake shape around 130 m.Source: Figure 14 · p. 32, §4 · pp. 32–33
  • At 3 m, CFM captures street-canyon gust curvature and most isolated-building patterns, but underestimates some magnitudes and has lee-side errors up to 1.5 m/s.The isolated-building southeast face is slightly overpredicted.Source: Figure 15 · p. 33, §4 · p. 33
  • Across three characteristic volumes, gust agreement is strong overall, though isolated-building flow is harder to capture because turbulence is advected from upstream.The isolated case also has smaller gusts than the other locations.Source: Figure 16 · p. 34, §4 · p. 34
  • Future work calls for continuous spatiotemporal generation, geometry embeddings that generalize across complex layouts, dynamic wind conditioning, and tests on additional CFD datasets.The authors also identify structural engineering as a potential application for further study.Source: §4 · p. 34

5. Summary - Conclusion

The paper applies CFM to generate instantaneous 3D urban microclimate fields and reports agreement with reference data across flow and turbulence statistics. Its conclusion presents the approach as a computationally efficient route to stochastic field generation and wind-gust estimation.

5. Summary - Conclusion

  • Validation reports 2.99% wind and 1.77% temperature mean-flow NRMSE, 7.17% and 8.84% standard-deviation NRMSE, and 7% TKE NRMSE.PDFs, including non-Gaussian tails and bimodal distributions, are also reported as captured with good accuracy.Source: §5 · p. 35
  • The CFM framework generates physically plausible instantaneous velocity and temperature snapshots from building geometry and mean flow, reproducing the stochastic character of LES data.It was trained on 18 urban configurations without turbulence statistics as input.Source: §5 · p. 35
  • The authors present CFM as a computationally efficient path to stochastic urban microclimate fields, probabilistic comfort assessment, and wind-gust estimation at a fraction of full-LES campaign cost.Source: §5 · p. 35

Related papers

Loading 2610.10430v1…