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

arXiv:2610.10430 · physics.flu-dyn, cs.CE, cs.LG, physics.ao-ph · Submitted 2026-10-07 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Conditional Flow Matching for Generation of 3D Multi-variable Instantaneous Urban Microclimate Fields".

Jane: The gist Rapid and accurate prediction of urban wind and temperature fields is important for urban microclimate design and climate adaptation (Page 1).

Tom: First, who's behind it and why it matters.

Paper summary: Jane: So we're diving into this paper called "Conditional Flow Matching for Generation of three dee Multi-variable Instantaneous Urban Microclimate Fields," and it seems like they've developed a really fast generative system <ref:2610.10430#pg1>. What does that mean for us who are just listening right now?

Tom: It means we can get high-resolution, three-dimensional snapshots of wind and temperature fields in seconds, which is crucial for things like urban design and climate adaptation <ref:2610.10430#pg1>.

Lu: They are moving away from those slow, deterministic point predictions that just ignore the turbulence. Instead, they’re learning the distribution of what a real turbulent flow looks like <ref:2610.10430#pg3>.

Jane: The paper uses Conditional Flow Matching to guide a generative model using building geometry and mean flow data to create these fields, and they achieved an RMSE of around one point nine two five percent when comparing against the mean flow field <ref:2610.10430#pg24>.

Tom: That accuracy proves the generated ensemble is physically anchored to the input, capturing structures like building wakes and temperature plumes without just being random noise <ref:2610.10430#pg19>, even showing complex nonGaussian distributions in temperature data <ref:2610.10430#pg35>.

Lu: They also addressed gust prediction, showing good correlation with CityFFD data for gusts around three targeted buildings <ref:2610.10430#pg32>.

Jane: So the implication is that we can assess urban comfort and wind conditions probabilistically, which is a big step up from just knowing a single number.

Tom: It’s about making probabilistic comfort assessment possible at a fraction of the computational cost of running full LES simulations <ref:2610.10430#pg35>.

Lu: The authors themselves noted that while the model does reproduce structures well, it has these limitations: it might act like a low-pass filter, underestimating peak fluctuation intensities in high-fluctuation zones <ref:2610.10430#pg30>, and they also admit that the generated patches aren't perfectly continuous across boundaries <ref:2610.10430#pg12>.

Jane: So the path forward involves using recurrent mechanisms or temporal attention layers to generate continuous flow series, and using Graph Neural Networks for better spatial efficiency with city layouts <ref:2610.10430#pg32>.

Tom: They say initial results on wind gusts give confidence to try this CFM framework for structural engineering applications too <ref:2610.10430#pg35>.

Lu: That’s the gist of it—a computationally efficient way to get LES resolution for stochastic urban microclimate fields.

Conclusion: Jane: Building on what we just heard, this paper, "Conditional Flow Matching for Generation of three dee Multi-variable Instantaneous Urban Microclimate Fields," shows how they’ve built a system that can generate detailed, real-time wind and temperature fields in seconds <ref:2610.10430#pg1>.

Tom: Exactly. We're talking about moving past the slow, heavy computational fluid dynamics work and making it fast enough for something like real-time urban design or even just checking if a street is going to be windy <ref:2610.10430#pg2>.

Lu: The core idea here is using conditional flow matching to guide a generative model, using building geometry and the average wind as a roadmap for creating these instantaneous snapshots <ref:2610.10430#pg8>.

Meng: From an engineering standpoint, I’m interested in how they handle the conditioning inputs because that’s where most of the complexity lies when you try to deploy something like this <ref:2610.10430#pg9>.

Lalam: My analysis shows that the authors are using a three-dimensional convolution U-Net architecture with self-attention layers to capture those long-range dependencies in the flow field data <ref:2610.10430#pg9>.

Tom: And they’re training it on CityFFD data, which is a GPU accelerated CFD solver employing a semi-Lagrangian approach, so the training data itself is super realistic <ref:2610.10430#pg5>.

Jane: So what does this mean for the people listening right now? It means we can get probabilistic comfort assessments for cities much faster than before.

Tom: Exactly. We’re moving from just getting one static answer to generating an ensemble of snapshots, which helps us understand how much the wind or temperature might fluctuate at a specific spot <ref:2610.10430#pg22>.

Lu: The results show that they reproduce the mean flow field with high fidelity, and crucially, they capture those complex turbulent features like building wakes and recirculation zones without any visible gaps between patches <ref:2610.10430#pg19>.

Meng: But I’ve seen papers before where generating patches independently leads to those nasty discontinuities at the boundaries; how did they manage that spatial continuity?

Jane: That’s a fair question. They tackled that by using a shared-noise initialization during patch generation, which kept the flow structure smooth across the whole domain <ref:2610.10430#pg12>.

Lalam: The paper confirms that generating one instantaneous snapshot takes about ten to thirteen seconds, meaning you can get thirty-five snapshots in less than eight minutes <ref:2610.10430#pg18>.

Tom: And when we look at the statistics, the normalized root mean square error across all three velocity components was around one point nine two five percent against the average flow field <ref:2610.10430#pg24>. That’s a solid anchor on how physically correct these results are.

Jane: So, while it’s fast and captures turbulence well—even nonGaussian temperature distributions—there are still some caveats, like it sometimes acting like a low-pass filter in really high-fluctuation areas <ref:2610.10430#pg30>.

Lu: The authors themselves pointed out that the patch-based generation isn't perfectly spatially coherent when you assemble them naively, which is why they’re looking toward future work involving recurrent mechanisms or Graph Neural Networks for better geometric embedding <ref:2610.10430#pg32>.

Meng: So the immediate implication is that we can use this for probabilistic wind and temperature comfort assessment at a much lower cost than running a full large eddy simulation campaign <ref:2610.10430#pg35>.

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

Concordia University · McGill University

physics.flu-dyn, cs.CE, cs.LG, physics.ao-ph

Submitted: 2026-10-07

Updated: 2026-10-07

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 84/100

The gist: The gist Rapid and accurate prediction of urban wind and temperature fields is important for urban microclimate design and climate adaptation (Page 1).

Key concepts

Large Eddy Simulation (LES): A high-fidelity CFD approach used to model urban microclimates by explicitly resolving turbulent fluctuations. While accurate, LES is computationally too expensive for real-time iterative design workflows, requiring hours or days of CPU time for a single scenario.
Conditional Flow Matching (CFM): A generative framework that uses building geometry and mean flow as conditions to guide the generation of plausible 3D velocity and temperature fields. It employs Optimal Transport to learn an efficient vector field that transforms random noise into the desired turbulent data distribution.
Optimal Transport (OT): A mathematical technique used in CFM to learn a retractable, direct intermediate vector field. This field guides sample points from a source of Gaussian noise toward the target data distribution, making the generation process more efficient and effective at capturing complex flow structures.
Conditioning Inputs: The model takes 15 channels of input, including building binary masks, Signed Distance Fields (SDFs) for building proximity, time-averaged steady-state flows (u, v, w, T), and roof height maps. These inputs guide the generation process to ensure the output fields are physically consistent with the specific urban environment.

Terminology

Flow Matching — 1× in this paper · explained in 15 episodes

Summary

The gist Rapid and accurate prediction of urban wind and temperature fields is important for urban microclimate design and climate adaptation (Page 1).

Problem Statement

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 (Page 1). Existing regressive data-driven models offers quick outputs, but they produce only deterministic point predictions that inherently fail to represent turbulent stochasticity (Page 1). These regressive frameworks inherently struggle to capture the transient, multi-scale nature of turbulent flows because they are typically optimized to minimize a point-to-point loss, tending to predict the conditional expectation and effectively averaging out the stochastic fluctuations inherent in turbulence (Page 2).

Proposed Solution: Conditional Flow Matching (CFM)

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 (Page 1). CFM introduces Ordinary Differential Equation (ODE) for calculating sampling timesteps, along with Optimal Transport (OT) to learn a retractable, direct, and highly efficient intermediate vector field that will lead the sample points from the source gaussian noise to the target data distribution (Page 8). The CFM objective is formulated as L2 loss between the parameterized vector field trained with a NN and the ground truth vector field u t(xz), where z is a conditioning variable (Page 8).

Methodology Details

The methodology involves several key components to address computational and spatial challenges:

Training Data: The training data for the generative model are produced using CityFFD, a GPU-accelerated CFD solver that employs a semi-Lagrangian approach coupled with a fractional step method (Page 5). Building cluster geometries for all cities used are selected around the world and were downloaded from the CityFFD platform (Page 5). The dataset is composed of 20 diverse and realistic urban morphologies for which 500 instantaneous flow field 3D snapshots were extracted (Page 6).
Model Architecture: The backbone architecture is a 3D convolution U-Net with a base channel width of 64 and channel multipliers of (1, 2, 4, 4) for four levels of channels (Page 9). Self-attention is added at the bottom layers with resolutions of 8 and 16 to ensure general global context modeling and long-range dependency capture in the flow field (Page 9).
Conditioning Inputs: The model inputs in total have 15 channels, involving noisy states, conditions, and target states of instantaneous flow field snapshots (Page 9). These include a building binary mask (1 ch), a Signed Distance Field (SDF) encoding the distance from each voxel to the nearest building surface (1 ch), the time-averaged steady-state flow field from CityFFD (4 ch: u, v, w, T), and a scalar roof height map (1 ch) (Page 9).
Patch-based Generation: To overcome 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 (Page 1). The study adopts a strategy of localized patches to address the hurdle without the needs of a pretrained encoder-decoder neural network to convert the full domain into latent space (Page 12).

Validation and Results

The validation framework is designed to validate the generative model across a robust system of statistical stringency and errors for some localized metric (Page 16). The ensemble mean of the 35 generated snapshots is compared against the CityFFD time-averaged mean flow field to verify that the generative model reproduces the conditioning mean flow with sufficient accuracy (Page 22). The overall normalized RMSE across the three velocity components is around 1.925%, confirming that the generated ensemble is physically anchored to the conditioning mean flow input without systematic drifting (Page 24).

Key Findings

Instantaneous Flow Generation: Generating an instantaneous 3D urban microclimate snapshot takes around 10 to 13 seconds, making the total time for generating a statistically converged ensemble of 35 instantaneous snapshots in less than 8 minutes (Page 18). The model produces physically structured instantaneous fields that exhibit realistic turbulent features like building wakes, street canyon vortices and various sizes of recirculation zones without visible discontinuities at patch boundaries (Page 19).
Flow Field Statistics: The model successfully reproduces the mean flow field with high fidelity, capturing the building aerodynamic and thermal plume structures (Page 23). The standard deviation between the generated snapshots for testing cities are obtained to quantify the amount of fluctuations to understand some of the regional strength or weaknesses of the generated results (Page 14).
Turbulence Generation: The CFM effectively replicates high-energy hotspots generated by shear layers along windward corners and captures sharp, discontinuous TKE gradients at the building-fluid interface (Page 11). The model captures complex, nonGaussian distributions including bimodal peaks in temperature PDFs, indicating that the flow-matching objective can learn complex, nonGaussian scalar transport phenomena where sharp thermal gradients are present (Page 35).
Wind Gust Prediction: Wind gusts predictions by CFM, based on the quasi-steady assumption around a region of three characteristic targeted buildings in the urban area were also reported and correlated very well with CityFFD data (Page 32).

Limitations and Future Work

The model exhibits a slight tendency to act as a low-pass filter, occasionally underestimating peak fluctuation intensities in high-fluctuation zones (Page 30). The practical limitation of patch-based generation is that independently generated patches are not spatially coherent and naively assembling them would produce visible discontinuities at patch boundaries (Page 12). Future research should prioritize the transition from static snapshot synthesis to the generation of continuous spatiotemporal wind flow series by integrating recurrent mechanisms or temporal attention layers (Page 32). The adoption of Graph Neural Networks (GNNs) for geometric embedding represents a critical next step to generalize across complex, non-orthogonal city layouts with high spatial efficiency (Page 32). The initial results regarding the gusts present in Section 4 provide confidence to try CFM for structural engineering applications as well (Page 35).

The CFM framework provides an effective and computationally efficient path to LES resolution stochastic urban microclimate fields, enabling probabilistic wind and temperature comfort assessment, and wind gust estimation at a fraction of the cost of full LES campaigns (Page 35).

--- Page 1 ---

Conditional Flow Matching for Generation of 3D Multi-variable Instantaneous Urban Microclimate Fields Peng Liu, Shaoxiang Qin, Theodore Potsis, Lili Ji, Dingyang Geng, and Liangzhu Leon Wang Conditional Flow Matching for Generation of 3D Multi-variable Instantaneous Urban Microclimate Fields

--- Page 2 ---

1 Introduction With the global urban population projected to reach 70% by 2050, the 21st century is characterized by an unprecedented expansion of the artificial land cover (Toparlar et. al, 2017) Urban microclimate governs a range of phenomena that directly affect the quality of life in cities, from pedestrian wind comfort and natural ventilation to pollutant dispersion, building aerodynamics, and outdoor thermal comfort (Yang et al., 2023) Computational Fluid Dynamics (CFD) approaches like Reynolds Averaged Navier Stokes (RANS) and Large Eddy Simulation (LES) are the most capable tools to model urban microclimates. Regarding RANS, the approximation of fluctuations that are needed for more detailed studies is insufficient, and in combination with the increasing computational resources a transition to LES has been seen in the current state-of-the-art (Potsis et al., 2023) However, besides High-Performance Computing accessibility, LES remains computationally prohibitive for real-time iterative design workflows: a single urban scenario on a domain of several hundred million grid cells requires hours to days of CPU time and producing the necessary for probabilistic analysis multiplies this cost further (Blocken, 2018; Yang, 2015)

--- Page 3 ---

Generative models offer a promising path to tackle such limitations for these regressive neural networks, such as providing manifolds of turbulent uncertainty (Du et al., 2024) Unlike deterministic regression, probabilistic generative models are designed to learn the underlying data distribution, which benefit the tasking aiming to capture the stochastic nature of turbulence (Li et al., 2023)

--- Page 4 ---

In generation tasks for large 3D urban instantaneous CFD flow field, there exist a long theoretical conflict between pixel-space direction generation with computational resource limitations, especially GPU ram for training (Du et al., 2024) To tackle such trade-off, recently papers in the machine learning field for urban microclimate studies has resorted to use overlapping patches to prevent memory outage and maintain boundary continuity for global flow field (Arakawa et al., 2023)

--- Page 5 ---

2.1 Numerical simulations The training data for the generative model are produced using CityFFD, a GPU-accelerated CFD solver that addresses the computational bottleneck of traditional methods by employing a semi-Lagrangian approach coupled with a fractional step method (Katal et al.

Improvements for AI systems

Improved AI system can generate physically plausible instantaneous realizations of 3D urban microclimate fields, which is superior to deterministic models that inherently fail to represent turbulent stochasticity.
The system can perform probabilistic assessments, enabling users to compute the probability of exceeding a discomfort threshold at each point in the domain, which moves beyond binary pass/fail classifications.
It can provide a tail behavior of the velocity distribution, which is critical for determining peak dynamic loads that govern design, something deterministic surrogates cannot support.
The system can generate wind gust predictions by CFM, based on the quasi-steady assumption, allowing for wind gust estimation in critical engineering parameters like Davenport's Peak Factor Method.
The system can be integrated into iterative urban design workflows at a reduced computational cost, enabling parametric studies that currently rely on RANS or deterministic surrogates to capture turbulence phenomena.