From 3d2ab03e718c3a8d036f78ad6402b5a8f77441b4 Mon Sep 17 00:00:00 2001 From: Isacco Date: Mon, 13 Jul 2026 10:58:21 +0200 Subject: [PATCH 1/6] zou's dataset for combustion GPR+HOSVD --- _databases/combustion_zou.md | 15 +++++++++++++++ 1 file changed, 15 insertions(+) create mode 100644 _databases/combustion_zou.md diff --git a/_databases/combustion_zou.md b/_databases/combustion_zou.md new file mode 100644 index 000000000000..28a6d1a92f98 --- /dev/null +++ b/_databases/combustion_zou.md @@ -0,0 +1,15 @@ +--- +layout: post +title: "Combustion data" +type: "Numerical / Combustion CFD" +tldr: "RANS CFD dataset of the DLR methane/hydrogen/nitrogen turbulent jet diffusion flame for reduced-order and data-driven modelling" +order: 4 +--- + +### Data index: + * [Zou's dataset](#zou-dataset) + +## Zou's dataset +This database contains RANS CFD simulations of the DLR turbulent jet diffusion flame, a reference configuration for turbulent non-premixed combustion consisting of a methane/hydrogen/nitrogen fuel jet issuing into a coflowing air stream. Simulations were run in OpenFOAM (`reactingFoam`) using the standard k-ε turbulence model and the Eddy Dissipation Concept (EDC) for finite-rate chemistry, and were validated against experimental temperature profiles. The dataset comprises 100 CFD cases generated by varying the fuel-jet Reynolds number (11,000-20,000) and the hydrogen mass fraction (4%-22%), with velocity, temperature, pressure, and species mass fraction fields sampled from each case and organized in tensor form for reduced-order and machine-learning modelling. + +Database available *coming soon*. From 56501ff0ec7b5d5b1e90307a7d88411357b9a0a5 Mon Sep 17 00:00:00 2001 From: Isacco Date: Mon, 13 Jul 2026 12:19:38 +0200 Subject: [PATCH 2/6] add BLASTNet dataset section --- _databases/combustion_zou.md | 14 ++++++++++---- 1 file changed, 10 insertions(+), 4 deletions(-) diff --git a/_databases/combustion_zou.md b/_databases/combustion_zou.md index 28a6d1a92f98..c6d35c4bd8a1 100644 --- a/_databases/combustion_zou.md +++ b/_databases/combustion_zou.md @@ -2,14 +2,20 @@ layout: post title: "Combustion data" type: "Numerical / Combustion CFD" -tldr: "RANS CFD dataset of the DLR methane/hydrogen/nitrogen turbulent jet diffusion flame for reduced-order and data-driven modelling" +tldr: "Combustion data used for research in reactive flows order: 4 --- ### Data index: - * [Zou's dataset](#zou-dataset) + * [DLR turbulent jet diffusion methane-hydrogen flame](#zou-dataset) + * [BLASTNet](#blastnet-dataset) -## Zou's dataset -This database contains RANS CFD simulations of the DLR turbulent jet diffusion flame, a reference configuration for turbulent non-premixed combustion consisting of a methane/hydrogen/nitrogen fuel jet issuing into a coflowing air stream. Simulations were run in OpenFOAM (`reactingFoam`) using the standard k-ε turbulence model and the Eddy Dissipation Concept (EDC) for finite-rate chemistry, and were validated against experimental temperature profiles. The dataset comprises 100 CFD cases generated by varying the fuel-jet Reynolds number (11,000-20,000) and the hydrogen mass fraction (4%-22%), with velocity, temperature, pressure, and species mass fraction fields sampled from each case and organized in tensor form for reduced-order and machine-learning modelling. +## DLR turbulent jet diffusion methane-hydrogen flame +This database contains RANS CFD simulations of the DLR turbulent jet diffusion flame, a reference configuration for turbulent non-premixed combustion consisting of a methane/hydrogen/nitrogen fuel jet issuing into a coflowing air stream. Simulations were run in OpenFOAM (`reactingFoam`) using the standard k-ε turbulence model and the Eddy Dissipation Concept (EDC) for finite-rate chemistry, and were validated against experimental temperature profiles. The dataset comprises 400 CFD cases generated by varying the fuel-jet Reynolds number (11,000-20,000) and the hydrogen mass fraction (4%-22%), with velocity, temperature, pressure, and species fields sampled from each case and organized in tensor form for reduced-order and machine-learning modelling. Database available *coming soon*. + +## BLASTNet +BLASTNet (Bearable Large Accessible Scientific Training Network-of-datasets) is an open, community-maintained collection of high-fidelity simulation data for fluid mechanics and combustion research, comprising several terabytes of turbulent and reacting-flow datasets used for turbulence closure modelling, spatio-temporal prediction, and inverse modelling. Data and pre-trained model weights are hosted on Kaggle, with accompanying code distributed via GitHub. + +[[Link to BLASTNet](https://blastnet.github.io/)] From d7575f2517efcc1af6a815952a59abe43bfee2af Mon Sep 17 00:00:00 2001 From: ariol Date: Thu, 23 Jul 2026 11:29:06 +0200 Subject: [PATCH 3/6] AIJ datasets added in software/datasets/urbanFlowData --- _databases/urban_canonical_9_buildings.md | 24 + .../Accelerate CFD/Accelerate_RANS.md | 520 +++++++++--------- 2 files changed, 284 insertions(+), 260 deletions(-) diff --git a/_databases/urban_canonical_9_buildings.md b/_databases/urban_canonical_9_buildings.md index 385e874ed6c4..8422927c8954 100644 --- a/_databases/urban_canonical_9_buildings.md +++ b/_databases/urban_canonical_9_buildings.md @@ -9,6 +9,10 @@ order: 3 ### Data index: * [9 buildings canonical configuration - Equispaced Array](#urban-9block-equispaced) * [9 buildings canonical configuration - Variable Height and Distance Array](#urban-9block-variable) + * [AIJ Case E - Niigata city district](#niigara-city-district) + * [AIJ Case K - Regular 9×9 cubic array](#9x9-cubic-array) + * [AIJ Case L - Regular 14×9 cubic array](#14x9-cubic-array) + * [AIJ Case M - TPU university campus](#tpu-university-campus) ## 9 buildings canonical configuration - Equispaced Array This database contains data from three-dimensional CFD simulations of turbulent flow around a 3x3 array of 9 cubic buildings, each 0.2 m x 0.2 m x 0.2 m (H = 0.2 m), @@ -25,3 +29,23 @@ Simulations were run varying the Reynolds number, the wind direction, and the em The dataset comprises 243 simulation cases, provided in .vtu (native solver mesh) and .npy (interpolated onto an equispaced grid) formats. Database available *coming soon*. + +## Urban LES database — AIJ Case E (Niigata city district) +This database contains data from a three-dimensional large-eddy simulation of turbulent flow over a real urban district in Niigata (Japan), reproducing the geometry of the AIJ Case E wind tunnel benchmark at full scale (footprint of about 395 m × 395 m, tallest building H = 60 m). Unlike the benchmark, which reports time-averaged measurements at sparse probe points, the simulation was run with OpenFOAM (WALE subgrid-scale model, synthetic-turbulence inlet from the measured profile) and validated against those points, providing the full time-resolved fields. The dataset comprises 497 snapshots of pressure and the three velocity components, provided in .vtu (native solver mesh) and .npy (equispaced grid of 198 × 198 × 21 points at Δxy = 2 m, Δz = 5 m, with building mask) formats. + +[[Link to EchoNet-Dynamic](https://www.aij.or.jp/%20jpn/publish/cfdguide/index_e.htm)] + +## Urban LES database — AIJ Case K (regular 9×9 cubic array) +This database contains data from a three-dimensional large-eddy simulation of turbulent flow through a regular array of 81 cubic buildings in 9 rows × 9 columns, of side H = 0.06 m and pitch 2H, reproducing the AIJ Case K (ArraysC) wind tunnel benchmark at model scale (U_H = 2.33 m/s, Re ≈ 9800). Unlike the benchmark, which reports time-averaged velocity and turbulent kinetic energy at 203 probe points, the simulation was run with OpenFOAM (WALE subgrid-scale model, synthetic-turbulence inlet with the measured anisotropic Reynolds stresses), validated against those points, and rescaled by a factor of 400 (H = 24 m, pitch 48 m). The dataset comprises 299 snapshots of pressure and the three velocity components, provided in .vtu (native solver mesh) and .npy (equispaced grid of 205 × 205 × 21 points at Δxy = 2 m, Δz = 5 m, with building mask) formats. + +[[Link to EchoNet-Dynamic](https://www.aij.or.jp/%20jpn/publish/cfdguide/index_e.htm)] + +## Urban LES database — AIJ Case L (regular 14×9 cubic array) +This database contains data from a three-dimensional large-eddy simulation of turbulent flow through a regular array of 126 cubic buildings in 14 rows × 9 columns, of side H = 0.06 m and pitch 2H, reproducing the AIJ Case L (ArraysCT) wind tunnel benchmark at model scale (U_H = 1.34 m/s). Unlike the benchmark, which was conducted under weakly non-isothermal conditions (Rib ≈ −0.04) and reports time-averaged point measurements, the simulation was run isothermally with OpenFOAM (WALE subgrid-scale model, synthetic-turbulence inlet from the measured profile), validated against the benchmark velocities, and rescaled by a factor of 400 (H = 24 m, pitch 48 m). The dataset comprises 299 snapshots of pressure and the three velocity components, provided in .vtu (native solver mesh) and .npy (equispaced grid of 325 × 205 × 21 points at Δxy = 2 m, Δz = 5 m, with building mask) formats. + +[[Link to EchoNet-Dynamic](https://www.aij.or.jp/%20jpn/publish/cfdguide/index_e.htm)] + +## Urban LES database — AIJ Case M (TPU university campus) +This database contains data from a three-dimensional large-eddy simulation of turbulent flow over a real university campus in Atsugi (Japan), reproducing the geometry of the AIJ Case M (TPU) wind tunnel benchmark at full scale (domain of about 1080 m × 720 m, tallest building 65 m, terrain elevation included). Unlike the benchmark, which reports time-averaged measurements at sparse points for several wind directions, the simulation was run for a single direction (180°, U_R = 4.08 m/s at H_R = 41 m) with OpenFOAM (WALE subgrid-scale model, synthetic-turbulence inlet from the measured profile) and validated against those measurements. The dataset comprises 450 snapshots of pressure and the three velocity components, provided in .vtu (native solver mesh) and .npy (equispaced grid of 329 × 292 × 21 points at Δxy = 2 m, Δz = 5 m, with building mask) formats. + +[[Link to EchoNet-Dynamic](https://www.aij.or.jp/%20jpn/publish/cfdguide/index_e.htm)] \ No newline at end of file diff --git a/_research/cfd-simulations/Accelerate CFD/Accelerate_RANS.md b/_research/cfd-simulations/Accelerate CFD/Accelerate_RANS.md index 2076c6054551..f5fe59bb415b 100644 --- a/_research/cfd-simulations/Accelerate CFD/Accelerate_RANS.md +++ b/_research/cfd-simulations/Accelerate CFD/Accelerate_RANS.md @@ -1,260 +1,260 @@ ---- -layout: page -category: "Adaptive Methodologies" -topic: "RANS acceleration" -tldr: "A divergence-aware adaptive CFD-surrogate framework that alternates between fast POD-DL forecasting and targeted OpenFOAM recalls for robust long-horizon prediction of unsteady flows." -title: "ROMIA: A methodology to accelerate RANS" -author: "Ángel Alonso, Miguel Rios" ---- - - -# Modal Prediction Pipeline for Accelerating Urban Flow CFD Simulations - -RANS simulations of urban or industrial flows on meshes of several million cells can require thousands of iterations to converge, with computation times of hours to days even on parallel clusters. A large fraction of this cost is concentrated in the final convergence stage, where the dominant flow structure is already established and the solver is simply relaxing towards the steady-state attractor. This work presents **ROMIA (Reduced Order Model for Industrial Acceleration)**, a non-intrusive modal prediction pipeline that learns a reduced-order model of the flow dynamics from intermediate simulation snapshots, extrapolates the asymptotic state, and reinjects it into the solver as an initial condition — bypassing the computationally expensive tail of the simulation. - -The pipeline is developed by the [ModelFlows group](https://modelflows.github.io/modelflowsapp/) at Universidad Politécnica de Madrid. This study constitutes the **first validation of ROMIA**, proving the value of the approximation on industrial and urban cases. - ---- - -# Methodology - -## Pipeline Overview - -ROMIA operates as a non-intrusive layer on top of OpenFOAM, requiring no modification of the solver. It chains four sequential stages during a running simulation: - - -

-ROMIA pipeline -

-*Figure 1. Schematic of the ROMIA pipeline. Snapshots captured during the transient phase are compressed spatially by iPOD, modelled temporally by HODMD, and the predicted asymptotic state is reconstructed and reinjected into the solver when the trigger criterion is satisfied.* - -## Spatial Reduction: Incremental POD (iPOD) - -At each snapshot interval, the current flow field, that lives in a space of dimensions N on the orders of tens of millions of parameters, is projected onto a low-dimensional basis maintained by Incremental Proper Orthogonal Decomposition (iPOD). The basis is updated online as new snapshots arrive, without recomputing the full SVD from scratch. The number of retained modes is controlled by an energy threshold ε₁, defined as the fraction of modal energy discarded relative to the total. This reduces the high-dimensional state to a compact set of temporal coefficients a(t) ∈ ℝʳ, where r ≪ N. - -## Temporal Modelling: HODMD - -Higher-Order Dynamic Mode Decomposition (HODMD) is applied to the series of reduced coefficients a(t). By constructing augmented snapshot matrices with a time-delay parameter d, HODMD identifies the spectral content of the dynamics even when the number of physically distinct frequencies exceeds the spatial rank of the reduced basis. Each extracted mode carries a growth rate δₘ and a frequency ωₘ. - -For stationary attractors, only modes with growth rate close to zero (|λⱼ| − 1 ≤ ε_δ) and frequency close to zero (|ℑ(ωⱼ)| ≤ ε_φ) are retained. The asymptotic state is predicted as the superposition of these stationary modes: - -**a**_∞ = Re( Σⱼ∈S ψⱼ bⱼ ) - -where S denotes the set of stationary modes, ψⱼ are the modal shapes and bⱼ their amplitudes. - -## Automatic Trigger - -The pipeline monitors prediction quality in real time through three criteria: spectral stability of the identified stationary modes, one-step prediction error against observed coefficients, and temporal stability of **a**_∞ between successive evaluations. When all criteria are simultaneously satisfied, the trigger authorises the jump: the simulation is stopped, the full physical field is reconstructed via - -**x**_pred = **x̄** + Φ **a**_∞ - -and reinjected into OpenFOAM as a restart condition. The solver then runs a short validation phase to confirm convergence. If the prediction is rejected, the simulation continues normally, so a failed prediction never corrupts the final result. - ---- - -# Validation Case I — Urban Flow - -## Geometry and Flow Conditions - -This validation case reproduces **AIJ Case C**, a canonical wind-tunnel benchmark of the Architectural Institute of Japan, consisting of a 3×3 array of nine square-section buildings of height H = 0.2 m with equal spacing H between them. The Reynolds number, based on the inflow velocity at building height U_H = 3.654 m/s and kinematic viscosity ν = 1.5 × 10⁻⁵ m²/s, is Re ≈ 4.9 × 10⁴. - - - -

-Geometry of the case -

- - - -## Computational Setup - -The CFD simulation was performed in **OpenFOAM v12** using the *simpleFoam* steady-state solver with a **RANS k–ε turbulence model**. The computational mesh is an unstructured polyhedral grid of **3,666,741 cells**, selected after a three-level convergence study (coarse / medium / fine) with two progressive refinement boxes around the building array. Boundary conditions include a stratified Dirichlet inlet profile reproducing the atmospheric boundary layer (ABL), no-slip conditions on all solid surfaces, and a zero-pressure outlet. - ---- - -# Parameter Sensitivity and Automatic Calibration - -ROMIA depends on five parameters that govern the different stages of the pipeline (ε₁, d, w, ε_δ, ε_φ). A key feature of the methodology is its ability to calibrate these parameters automatically during the simulation, without prior knowledge of the converged solution. To verify this capability and identify the operating ranges suitable for urban RANS flows, a systematic sensitivity analysis was conducted on the AIJ Case C configuration. - -Another key feature of ROMIA is its ability to trigger the reconstruction at the optimal moment, minimising the reconstruction error. This analysis also detects, with offline and previously simulated data, the optimal moment to trigger. - -## Sensitivity Analysis - -The following figures show the reconstruction error as a function of each parameter, with the remaining parameters held fixed at their baseline values. - - - -

- -

-*Figure 2. Sensitivity to the spatial energy threshold ε₁.* - -

- -

-*Figure 3. Sensitivity to the HODMD time-delay parameter d.* - -

- -

-*Figure 4. Relation between number of POD modes, snapshot of trigger, and reconstruction error.* - -

- -

-![best snapshot](07_tradeoff_cuantitativo_Uy.png) -*Figure 5. Optimal moment to trigger the reconstruction as a function of iteration number.* - - -The analysis of sensitivity shows that the best moment to trigger the prediction is around iteration 1250. The optimal number of retained modes lies between 10 and 15, with an optimal spatial energy threshold ε₁ = 10⁻⁴ and modal amplitude threshold ε₂ = 5 × 10⁻⁴. - -## Automatic Calibration Results - -The sensitivity analysis identifies the ranges within which accurate reconstruction is achievable. ROMIA's automatic calibration procedure, operating without access to the reference solution, selected the following parameters: - -| Parameter | Manual optimum | ROMIA selection | -|-----------|---------------|-----------------| -| Trigger iteration | 1250 | 1350 | -| d | 15 | 20 | - -The automatically selected values fall within the acceptable ranges identified by the sensitivity analysis, confirming that the calibration procedure generalises correctly to this flow regime. - -# Results — Urban Flow - -## Low-Dimensionality of the Flow - -The singular value spectrum of the iPOD basis confirms the low-dimensionality hypothesis underlying the ROM approach. With **r = 10 modes**, the pipeline retains **99.99% of the modal energy** of a state space of approximately 15 million parameters. This result demonstrates that, despite the geometric complexity of nine interacting bluff bodies, the transient evolution of the flow inhabits a subspace of very low dimension, validating the applicability of the ROM framework to this flow regime. - -## Prediction Accuracy - -The pipeline triggered at **iteration 1350** and predicted the converged state corresponding to **iteration 4000**, saving approximately **2200 solver iterations** and achieving a **speed-up of 2.2×** relative to the full simulation. - -Following the prediction of the converged state, ROMIA restarted the solver and ran 200 additional CFD iterations to validate the physical consistency of the reconstructed field and refine the solution. - -The figures below compare the ROMIA prediction against the CFD reference field on a horizontal slice at building height: - - -

- -

-*Figure 6. Comparasion between the reconstruction on Uy, p and the ground truth*. - - -Quantitative reconstruction errors are reported in the table below: - -| Field | L² relative error | -|-------|------------------| -| U_y (main flow direction) | ~3% | -| U_x (lateral) | ~15% | -| U_z (vertical) | ~15% | -| p (pressure) | ~15% | - -The dominant velocity component U_y is reconstructed with ~3% L² error. The higher relative errors in U_x, U_z and p are partly a metric artefact: these fields are nearly zero across most of the domain, concentrated in the inter-building recirculation zones, so small absolute errors in those regions produce disproportionately large relative values. Visual inspection of the field comparisons confirms that the spatial structure of all fields is correctly reproduced. - ---- - -# Validation Case II — Formula 1 Rear Wing - -The second validation case extends ROMIA to a fully industrial geometry: the rear wing of a Formula 1 single-seater, discretised on a mesh of **29,482,757 cells** — an order of magnitude larger than the urban case. The objective is to demonstrate that the pipeline scales to production-level HPC workloads and that the automatic calibration generalises across geometrically distinct flow regimes. - -## Geometry and Flow Conditions - -The computational domain includes the rear wing profile, its immediate wake, and the surrounding control volume. The inflow is horizontal with uniform velocity U∞ = 73 m/s, giving a Reynolds number of Re ≈ 3.5 × 10⁶ based on the mean wing chord. - - -

- -

-*Figure 7. Validation case: Formula 1 rear wing. Horizontal inflow at U∞ = 73 m/s.* - - - -## Computational Setup - -The CFD simulation was performed in **OpenFOAM 2506** using the *simpleFoam* steady-state solver with a **standard k–ε turbulence model**, executed in parallel on **72 processors**. ROMIA operated in managed mode with **10% spatial subsampling** (cell_volume_sqrt strategy), resulting in 3,537,930 subsampled cells. The pipeline was configured with a warmup phase of 100 iterations, followed by snapshot accumulation at writeInterval = 2 from iteration 102 onwards, producing 200 snapshots in batches of 5. - -## Results — Formula 1 Rear Wing - -### Low-Dimensionality of the Transient Flow - -The singular value spectrum exhibits a pronounced spectral separation: the first mode concentrates σ₁ = 4881.5, while σ₂ = 402.9 and σ₃ = 131.7. The ratio σ₁/σ₈₀ ≈ 5.77 × 10⁴ demonstrates extreme energy concentration in the leading modes, characteristic of external flows over aerodynamic bodies where a global displacement mode dominates the convergence trajectory. - -| Modes retained (r) | Cumulative modal energy | -|--------------------|------------------------| -| 2 | 99.85% | -| 5 | 99.989% | -| 10 | 99.999% | - -With only **5 POD modes on 10% of the cells**, ROMIA retains 99.989% of the modal energy from a state space of approximately 2.1 × 10⁷ degrees of freedom in the subsample. The spatial compression ratio is of the order N/r ≈ 2.6 × 10⁵. The iPOD basis stabilised operationally at **r = 80 modes** after approximately 80 snapshots (iteration ≈ 260). - -### Asymptotic State Identification - -The HODMD autocalibration procedure autonomously selected the optimal configuration **(d = 20, window = 180)**, operating over r = 80 POD modes. The fit extracted **4 dynamic modes**, all classified as stationary: - -| Mode | \|λ\| | Growth rate σ (iter⁻¹) | Frequency ω (rad/iter) | -|------|--------|------------------------|------------------------| -| 1 | 1.027 | 1.31 × 10⁻² | +5.29 × 10⁻³ | -| 2 | 1.027 | 1.31 × 10⁻² | −5.29 × 10⁻³ | -| 3 | 1.021 | 1.06 × 10⁻² | +1.50 × 10⁻² | -| 4 | 1.021 | 1.06 × 10⁻² | −1.50 × 10⁻² | - -No unstable or damped modes were detected. The oscillation frequencies (ω ∼ 10⁻²–10⁻³ rad/iteration) are compatible with quasi-stationary convergence towards the solver fixed point. - -### Field Reconstruction and Computational Speed-Up - -The trigger declared **READY at iteration 500**. Following reconstruction and restart, all residuals decay monotonically over 200 additional iterations, with no blow-up or numerical oscillations. The maximum residual at iteration 10200 was 1.06 × 10⁻⁴, comparable to the convergence level reached prior to restart. - -

- -

-*Figure 8. Velocity magnitude |U| Left: ROMIA reconstructed field (iter. 10000). Right: Ground truth.* - -

- -

-*Figure 9: Presion field Left: ROMIA reconstructed field (iter. 10000). Right: Ground truth.* - -

- -

-*Figure 10. L2 Error of U field* - - -

- -

-*Figure 11. L2 Error of presion field* - -The pipeline was evaluated against a baseline run without ROM intervention: - -| Configuration | Wall-clock time | -|---------------|----------------| -| Baseline (full simpleFoam) | 15 h 54 m | -| ROMIA (warmup + iPOD + trigger + restart) | 6 h 32 m | -| **Speed-up** | **2.49×** | - -ROMIA reduced the wall-clock time by a factor of approximately **2.5×**, saving roughly 60% of the total computation time on a 29.5-million-cell industrial mesh running on 72 parallel processors. - ---- - -# Summary - -The ROMIA pipeline has been validated on two geometrically distinct RANS flow regimes, demonstrating consistent behaviour across both cases: - -| | Urban flow (AIJ Case C) | F1 rear wing | -|---|---|---| -| Mesh size | 3.6 M cells | 29.5 M cells | -| POD modes (99.99% energy) | 10 | 5 | -| Trigger iteration | 1350 | 500 | -| Speed-up | 2.2× | 2.5× | -| L² error (dominant field) | ~3% (U_y) | — | - -Key findings across both cases: - -- The transient RANS dynamics inhabit a subspace of very low dimension relative to the full state space, confirming the low-dimensionality hypothesis in both flow regimes. -- HODMD autocalibration autonomously identifies the optimal time-delay and window parameters without manual intervention. -- The multi-layer trigger avoids premature reconstruction, waiting for sufficient snapshot accumulation and autocalibration validation before declaring READY. -- The reconstructed fields are accepted by the OpenFOAM solver without divergence, and residuals converge monotonically after restart in both cases. - ---- - +--- +layout: page +category: "Adaptive Methodologies" +topic: "RANS acceleration" +tldr: "A divergence-aware adaptive CFD-surrogate framework that alternates between fast POD-DL forecasting and targeted OpenFOAM recalls for robust long-horizon prediction of unsteady flows." +title: "ROMIA: A methodology to accelerate RANS" +author: "Ángel Alonso, Miguel Rios" +--- + + +# Modal Prediction Pipeline for Accelerating Urban Flow CFD Simulations + +RANS simulations of urban or industrial flows on meshes of several million cells can require thousands of iterations to converge, with computation times of hours to days even on parallel clusters. A large fraction of this cost is concentrated in the final convergence stage, where the dominant flow structure is already established and the solver is simply relaxing towards the steady-state attractor. This work presents **ROMIA (Reduced Order Model for Industrial Acceleration)**, a non-intrusive modal prediction pipeline that learns a reduced-order model of the flow dynamics from intermediate simulation snapshots, extrapolates the asymptotic state, and reinjects it into the solver as an initial condition — bypassing the computationally expensive tail of the simulation. + +The pipeline is developed by the [ModelFlows group](https://modelflows.github.io/modelflowsapp/) at Universidad Politécnica de Madrid. This study constitutes the **first validation of ROMIA**, proving the value of the approximation on industrial and urban cases. + +--- + +# Methodology + +## Pipeline Overview + +ROMIA operates as a non-intrusive layer on top of OpenFOAM, requiring no modification of the solver. It chains four sequential stages during a running simulation: + + +

+ROMIA pipeline +

+*Figure 1. Schematic of the ROMIA pipeline. Snapshots captured during the transient phase are compressed spatially by iPOD, modelled temporally by HODMD, and the predicted asymptotic state is reconstructed and reinjected into the solver when the trigger criterion is satisfied.* + +## Spatial Reduction: Incremental POD (iPOD) + +At each snapshot interval, the current flow field, that lives in a space of dimensions N on the orders of tens of millions of parameters, is projected onto a low-dimensional basis maintained by Incremental Proper Orthogonal Decomposition (iPOD). The basis is updated online as new snapshots arrive, without recomputing the full SVD from scratch. The number of retained modes is controlled by an energy threshold ε₁, defined as the fraction of modal energy discarded relative to the total. This reduces the high-dimensional state to a compact set of temporal coefficients a(t) ∈ ℝʳ, where r ≪ N. + +## Temporal Modelling: HODMD + +Higher-Order Dynamic Mode Decomposition (HODMD) is applied to the series of reduced coefficients a(t). By constructing augmented snapshot matrices with a time-delay parameter d, HODMD identifies the spectral content of the dynamics even when the number of physically distinct frequencies exceeds the spatial rank of the reduced basis. Each extracted mode carries a growth rate δₘ and a frequency ωₘ. + +For stationary attractors, only modes with growth rate close to zero (|λⱼ| − 1 ≤ ε_δ) and frequency close to zero (|ℑ(ωⱼ)| ≤ ε_φ) are retained. The asymptotic state is predicted as the superposition of these stationary modes: + +**a**_∞ = Re( Σⱼ∈S ψⱼ bⱼ ) + +where S denotes the set of stationary modes, ψⱼ are the modal shapes and bⱼ their amplitudes. + +## Automatic Trigger + +The pipeline monitors prediction quality in real time through three criteria: spectral stability of the identified stationary modes, one-step prediction error against observed coefficients, and temporal stability of **a**_∞ between successive evaluations. When all criteria are simultaneously satisfied, the trigger authorises the jump: the simulation is stopped, the full physical field is reconstructed via + +**x**_pred = **x̄** + Φ **a**_∞ + +and reinjected into OpenFOAM as a restart condition. The solver then runs a short validation phase to confirm convergence. If the prediction is rejected, the simulation continues normally, so a failed prediction never corrupts the final result. + +--- + +# Validation Case I — Urban Flow + +## Geometry and Flow Conditions + +This validation case reproduces **AIJ Case C**, a canonical wind-tunnel benchmark of the Architectural Institute of Japan, consisting of a 3×3 array of nine square-section buildings of height H = 0.2 m with equal spacing H between them. The Reynolds number, based on the inflow velocity at building height U_H = 3.654 m/s and kinematic viscosity ν = 1.5 × 10⁻⁵ m²/s, is Re ≈ 4.9 × 10⁴. + + + +

+Geometry of the case +

+ + + +## Computational Setup + +The CFD simulation was performed in **OpenFOAM v12** using the *simpleFoam* steady-state solver with a **RANS k–ε turbulence model**. The computational mesh is an unstructured polyhedral grid of **3,666,741 cells**, selected after a three-level convergence study (coarse / medium / fine) with two progressive refinement boxes around the building array. Boundary conditions include a stratified Dirichlet inlet profile reproducing the atmospheric boundary layer (ABL), no-slip conditions on all solid surfaces, and a zero-pressure outlet. + +--- + +# Parameter Sensitivity and Automatic Calibration + +ROMIA depends on five parameters that govern the different stages of the pipeline (ε₁, d, w, ε_δ, ε_φ). A key feature of the methodology is its ability to calibrate these parameters automatically during the simulation, without prior knowledge of the converged solution. To verify this capability and identify the operating ranges suitable for urban RANS flows, a systematic sensitivity analysis was conducted on the AIJ Case C configuration. + +Another key feature of ROMIA is its ability to trigger the reconstruction at the optimal moment, minimising the reconstruction error. This analysis also detects, with offline and previously simulated data, the optimal moment to trigger. + +## Sensitivity Analysis + +The following figures show the reconstruction error as a function of each parameter, with the remaining parameters held fixed at their baseline values. + + + +

+ +

+*Figure 2. Sensitivity to the spatial energy threshold ε₁.* + +

+ +

+*Figure 3. Sensitivity to the HODMD time-delay parameter d.* + +

+ +

+*Figure 4. Relation between number of POD modes, snapshot of trigger, and reconstruction error.* + +

+ +

+![best snapshot](07_tradeoff_cuantitativo_Uy.png) +*Figure 5. Optimal moment to trigger the reconstruction as a function of iteration number.* + + +The analysis of sensitivity shows that the best moment to trigger the prediction is around iteration 1250. The optimal number of retained modes lies between 10 and 15, with an optimal spatial energy threshold ε₁ = 10⁻⁴ and modal amplitude threshold ε₂ = 5 × 10⁻⁴. + +## Automatic Calibration Results + +The sensitivity analysis identifies the ranges within which accurate reconstruction is achievable. ROMIA's automatic calibration procedure, operating without access to the reference solution, selected the following parameters: + +| Parameter | Manual optimum | ROMIA selection | +|-----------|---------------|-----------------| +| Trigger iteration | 1250 | 1350 | +| d | 15 | 20 | + +The automatically selected values fall within the acceptable ranges identified by the sensitivity analysis, confirming that the calibration procedure generalises correctly to this flow regime. + +# Results — Urban Flow + +## Low-Dimensionality of the Flow + +The singular value spectrum of the iPOD basis confirms the low-dimensionality hypothesis underlying the ROM approach. With **r = 10 modes**, the pipeline retains **99.99% of the modal energy** of a state space of approximately 15 million parameters. This result demonstrates that, despite the geometric complexity of nine interacting bluff bodies, the transient evolution of the flow inhabits a subspace of very low dimension, validating the applicability of the ROM framework to this flow regime. + +## Prediction Accuracy + +The pipeline triggered at **iteration 1350** and predicted the converged state corresponding to **iteration 4000**, saving approximately **2200 solver iterations** and achieving a **speed-up of 2.2×** relative to the full simulation. + +Following the prediction of the converged state, ROMIA restarted the solver and ran 200 additional CFD iterations to validate the physical consistency of the reconstructed field and refine the solution. + +The figures below compare the ROMIA prediction against the CFD reference field on a horizontal slice at building height: + + +

+ +

+*Figure 6. Comparasion between the reconstruction on Uy, p and the ground truth*. + + +Quantitative reconstruction errors are reported in the table below: + +| Field | L² relative error | +|-------|------------------| +| U_y (main flow direction) | ~3% | +| U_x (lateral) | ~15% | +| U_z (vertical) | ~15% | +| p (pressure) | ~15% | + +The dominant velocity component U_y is reconstructed with ~3% L² error. The higher relative errors in U_x, U_z and p are partly a metric artefact: these fields are nearly zero across most of the domain, concentrated in the inter-building recirculation zones, so small absolute errors in those regions produce disproportionately large relative values. Visual inspection of the field comparisons confirms that the spatial structure of all fields is correctly reproduced. + +--- + +# Validation Case II — Formula 1 Rear Wing + +The second validation case extends ROMIA to a fully industrial geometry: the rear wing of a Formula 1 single-seater, discretised on a mesh of **29,482,757 cells** — an order of magnitude larger than the urban case. The objective is to demonstrate that the pipeline scales to production-level HPC workloads and that the automatic calibration generalises across geometrically distinct flow regimes. + +## Geometry and Flow Conditions + +The computational domain includes the rear wing profile, its immediate wake, and the surrounding control volume. The inflow is horizontal with uniform velocity U∞ = 73 m/s, giving a Reynolds number of Re ≈ 3.5 × 10⁶ based on the mean wing chord. + + +

+ +

+*Figure 7. Validation case: Formula 1 rear wing. Horizontal inflow at U∞ = 73 m/s.* + + + +## Computational Setup + +The CFD simulation was performed in **OpenFOAM 2506** using the *simpleFoam* steady-state solver with a **standard k–ε turbulence model**, executed in parallel on **72 processors**. ROMIA operated in managed mode with **10% spatial subsampling** (cell_volume_sqrt strategy), resulting in 3,537,930 subsampled cells. The pipeline was configured with a warmup phase of 100 iterations, followed by snapshot accumulation at writeInterval = 2 from iteration 102 onwards, producing 200 snapshots in batches of 5. + +## Results — Formula 1 Rear Wing + +### Low-Dimensionality of the Transient Flow + +The singular value spectrum exhibits a pronounced spectral separation: the first mode concentrates σ₁ = 4881.5, while σ₂ = 402.9 and σ₃ = 131.7. The ratio σ₁/σ₈₀ ≈ 5.77 × 10⁴ demonstrates extreme energy concentration in the leading modes, characteristic of external flows over aerodynamic bodies where a global displacement mode dominates the convergence trajectory. + +| Modes retained (r) | Cumulative modal energy | +|--------------------|------------------------| +| 2 | 99.85% | +| 5 | 99.989% | +| 10 | 99.999% | + +With only **5 POD modes on 10% of the cells**, ROMIA retains 99.989% of the modal energy from a state space of approximately 2.1 × 10⁷ degrees of freedom in the subsample. The spatial compression ratio is of the order N/r ≈ 2.6 × 10⁵. The iPOD basis stabilised operationally at **r = 80 modes** after approximately 80 snapshots (iteration ≈ 260). + +### Asymptotic State Identification + +The HODMD autocalibration procedure autonomously selected the optimal configuration **(d = 20, window = 180)**, operating over r = 80 POD modes. The fit extracted **4 dynamic modes**, all classified as stationary: + +| Mode | \|λ\| | Growth rate σ (iter⁻¹) | Frequency ω (rad/iter) | +|------|--------|------------------------|------------------------| +| 1 | 1.027 | 1.31 × 10⁻² | +5.29 × 10⁻³ | +| 2 | 1.027 | 1.31 × 10⁻² | −5.29 × 10⁻³ | +| 3 | 1.021 | 1.06 × 10⁻² | +1.50 × 10⁻² | +| 4 | 1.021 | 1.06 × 10⁻² | −1.50 × 10⁻² | + +No unstable or damped modes were detected. The oscillation frequencies (ω ∼ 10⁻²–10⁻³ rad/iteration) are compatible with quasi-stationary convergence towards the solver fixed point. + +### Field Reconstruction and Computational Speed-Up + +The trigger declared **READY at iteration 500**. Following reconstruction and restart, all residuals decay monotonically over 200 additional iterations, with no blow-up or numerical oscillations. The maximum residual at iteration 10200 was 1.06 × 10⁻⁴, comparable to the convergence level reached prior to restart. + +

+ +

+*Figure 8. Velocity magnitude |U| Left: ROMIA reconstructed field (iter. 10000). Right: Ground truth.* + +

+ +

+*Figure 9: Presion field Left: ROMIA reconstructed field (iter. 10000). Right: Ground truth.* + +

+ +

+*Figure 10. L2 Error of U field* + + +

+ +

+*Figure 11. L2 Error of presion field* + +The pipeline was evaluated against a baseline run without ROM intervention: + +| Configuration | Wall-clock time | +|---------------|----------------| +| Baseline (full simpleFoam) | 15 h 54 m | +| ROMIA (warmup + iPOD + trigger + restart) | 6 h 32 m | +| **Speed-up** | **2.49×** | + +ROMIA reduced the wall-clock time by a factor of approximately **2.5×**, saving roughly 60% of the total computation time on a 29.5-million-cell industrial mesh running on 72 parallel processors. + +--- + +# Summary + +The ROMIA pipeline has been validated on two geometrically distinct RANS flow regimes, demonstrating consistent behaviour across both cases: + +| | Urban flow (AIJ Case C) | F1 rear wing | +|---|---|---| +| Mesh size | 3.6 M cells | 29.5 M cells | +| POD modes (99.99% energy) | 10 | 5 | +| Trigger iteration | 1350 | 500 | +| Speed-up | 2.2× | 2.5× | +| L² error (dominant field) | ~3% (U_y) | — | + +Key findings across both cases: + +- The transient RANS dynamics inhabit a subspace of very low dimension relative to the full state space, confirming the low-dimensionality hypothesis in both flow regimes. +- HODMD autocalibration autonomously identifies the optimal time-delay and window parameters without manual intervention. +- The multi-layer trigger avoids premature reconstruction, waiting for sufficient snapshot accumulation and autocalibration validation before declaring READY. +- The reconstructed fields are accepted by the OpenFOAM solver without divergence, and residuals converge monotonically after restart in both cases. + +--- + From 00e73337954a3feb2f8ae1a9bc39cf31b63a19cc Mon Sep 17 00:00:00 2001 From: Paul Jeanney Date: Thu, 23 Jul 2026 15:01:39 +0200 Subject: [PATCH 4/6] Add Vallecas urban database information Added details about the Vallecas urban database, including simulation parameters and pollutant data. --- _databases/urban_canonical_9_buildings.md | 8 +++++++- 1 file changed, 7 insertions(+), 1 deletion(-) diff --git a/_databases/urban_canonical_9_buildings.md b/_databases/urban_canonical_9_buildings.md index 8422927c8954..6ed76db557f9 100644 --- a/_databases/urban_canonical_9_buildings.md +++ b/_databases/urban_canonical_9_buildings.md @@ -30,6 +30,12 @@ The dataset comprises 243 simulation cases, provided in .vtu (native solver mesh Database available *coming soon*. +## Vallecas real-urban large-scale configuration, real terrain elevation and building array + +This database contains data from three-dimensional CFD simulations of turbulent wind flow over the Vallecas district of Madrid, a real urban area with varying height and footprint set on real terrain elevation. The domain is a cylinder with a diameter of approximately 3 410 m and a vertical extent ranging between 317 m and 358 m. It is modelled with neutral atmospheric boundary layer conditions, computed with steady-state RANS and the $k$-$\varepsilon$ turbulence model (with simpleFoam). The ground is split into terrain, vegetation and water, each with its own surface roughness. Simulations were run varying the wind direction and reference wind speed at 5.5 m height. For each case the following fields are saved: U (mean velocity (m/s)), p (mean kinematic pressure (m²/s²)), k (turbulent kinetic energy (m²/s²)), epsilon (turbulent dissipation rate (m²/s³)), nut (turbulent eddy viscosity (m²/s)). In addition, traffic pollutants emitted from the nearby highways, with emission rates estimated from COPERT values, are transported as passive scalars, giving the concentration fields CO, NOx and PM (kg/m³). The dataset comprises 24 simulation cases, provided in .vtu format. + +Database available *coming soon*. + ## Urban LES database — AIJ Case E (Niigata city district) This database contains data from a three-dimensional large-eddy simulation of turbulent flow over a real urban district in Niigata (Japan), reproducing the geometry of the AIJ Case E wind tunnel benchmark at full scale (footprint of about 395 m × 395 m, tallest building H = 60 m). Unlike the benchmark, which reports time-averaged measurements at sparse probe points, the simulation was run with OpenFOAM (WALE subgrid-scale model, synthetic-turbulence inlet from the measured profile) and validated against those points, providing the full time-resolved fields. The dataset comprises 497 snapshots of pressure and the three velocity components, provided in .vtu (native solver mesh) and .npy (equispaced grid of 198 × 198 × 21 points at Δxy = 2 m, Δz = 5 m, with building mask) formats. @@ -48,4 +54,4 @@ This database contains data from a three-dimensional large-eddy simulation of tu ## Urban LES database — AIJ Case M (TPU university campus) This database contains data from a three-dimensional large-eddy simulation of turbulent flow over a real university campus in Atsugi (Japan), reproducing the geometry of the AIJ Case M (TPU) wind tunnel benchmark at full scale (domain of about 1080 m × 720 m, tallest building 65 m, terrain elevation included). Unlike the benchmark, which reports time-averaged measurements at sparse points for several wind directions, the simulation was run for a single direction (180°, U_R = 4.08 m/s at H_R = 41 m) with OpenFOAM (WALE subgrid-scale model, synthetic-turbulence inlet from the measured profile) and validated against those measurements. The dataset comprises 450 snapshots of pressure and the three velocity components, provided in .vtu (native solver mesh) and .npy (equispaced grid of 329 × 292 × 21 points at Δxy = 2 m, Δz = 5 m, with building mask) formats. -[[Link to EchoNet-Dynamic](https://www.aij.or.jp/%20jpn/publish/cfdguide/index_e.htm)] \ No newline at end of file +[[Link to EchoNet-Dynamic](https://www.aij.or.jp/%20jpn/publish/cfdguide/index_e.htm)] From 8f0370c9cb2dc5189b7e1c99b51cc6d214cc31ee Mon Sep 17 00:00:00 2001 From: Paul Jeanney Date: Thu, 23 Jul 2026 15:05:09 +0200 Subject: [PATCH 5/6] Update urban_canonical_9_buildings.md --- _databases/urban_canonical_9_buildings.md | 1 + 1 file changed, 1 insertion(+) diff --git a/_databases/urban_canonical_9_buildings.md b/_databases/urban_canonical_9_buildings.md index 6ed76db557f9..19d04f8d5032 100644 --- a/_databases/urban_canonical_9_buildings.md +++ b/_databases/urban_canonical_9_buildings.md @@ -9,6 +9,7 @@ order: 3 ### Data index: * [9 buildings canonical configuration - Equispaced Array](#urban-9block-equispaced) * [9 buildings canonical configuration - Variable Height and Distance Array](#urban-9block-variable) + * [Vallecas real-urban large-scale configuration, real terrain elevation and building array](#urban-vallecas) * [AIJ Case E - Niigata city district](#niigara-city-district) * [AIJ Case K - Regular 9×9 cubic array](#9x9-cubic-array) * [AIJ Case L - Regular 14×9 cubic array](#14x9-cubic-array) From b484934c272dc62cbe49dc6a0423fede203f64ac Mon Sep 17 00:00:00 2001 From: ariol Date: Thu, 23 Jul 2026 15:27:56 +0200 Subject: [PATCH 6/6] AIJ datasets added in software/datasets/urbanFlowData fixed --- _databases/urban_canonical_9_buildings.md | 36 ++++++----------------- 1 file changed, 9 insertions(+), 27 deletions(-) diff --git a/_databases/urban_canonical_9_buildings.md b/_databases/urban_canonical_9_buildings.md index 19d04f8d5032..06da5a3b6a4f 100644 --- a/_databases/urban_canonical_9_buildings.md +++ b/_databases/urban_canonical_9_buildings.md @@ -9,11 +9,9 @@ order: 3 ### Data index: * [9 buildings canonical configuration - Equispaced Array](#urban-9block-equispaced) * [9 buildings canonical configuration - Variable Height and Distance Array](#urban-9block-variable) - * [Vallecas real-urban large-scale configuration, real terrain elevation and building array](#urban-vallecas) - * [AIJ Case E - Niigata city district](#niigara-city-district) - * [AIJ Case K - Regular 9×9 cubic array](#9x9-cubic-array) - * [AIJ Case L - Regular 14×9 cubic array](#14x9-cubic-array) - * [AIJ Case M - TPU university campus](#tpu-university-campus) + * [AIJ - Benchmark geometris](#aij-data) + * [Urban wind tunnel database - EWTL Hamburg reference](#ewtl-data) + ## 9 buildings canonical configuration - Equispaced Array This database contains data from three-dimensional CFD simulations of turbulent flow around a 3x3 array of 9 cubic buildings, each 0.2 m x 0.2 m x 0.2 m (H = 0.2 m), @@ -31,28 +29,12 @@ The dataset comprises 243 simulation cases, provided in .vtu (native solver mesh Database available *coming soon*. -## Vallecas real-urban large-scale configuration, real terrain elevation and building array - -This database contains data from three-dimensional CFD simulations of turbulent wind flow over the Vallecas district of Madrid, a real urban area with varying height and footprint set on real terrain elevation. The domain is a cylinder with a diameter of approximately 3 410 m and a vertical extent ranging between 317 m and 358 m. It is modelled with neutral atmospheric boundary layer conditions, computed with steady-state RANS and the $k$-$\varepsilon$ turbulence model (with simpleFoam). The ground is split into terrain, vegetation and water, each with its own surface roughness. Simulations were run varying the wind direction and reference wind speed at 5.5 m height. For each case the following fields are saved: U (mean velocity (m/s)), p (mean kinematic pressure (m²/s²)), k (turbulent kinetic energy (m²/s²)), epsilon (turbulent dissipation rate (m²/s³)), nut (turbulent eddy viscosity (m²/s)). In addition, traffic pollutants emitted from the nearby highways, with emission rates estimated from COPERT values, are transported as passive scalars, giving the concentration fields CO, NOx and PM (kg/m³). The dataset comprises 24 simulation cases, provided in .vtu format. - -Database available *coming soon*. - -## Urban LES database — AIJ Case E (Niigata city district) -This database contains data from a three-dimensional large-eddy simulation of turbulent flow over a real urban district in Niigata (Japan), reproducing the geometry of the AIJ Case E wind tunnel benchmark at full scale (footprint of about 395 m × 395 m, tallest building H = 60 m). Unlike the benchmark, which reports time-averaged measurements at sparse probe points, the simulation was run with OpenFOAM (WALE subgrid-scale model, synthetic-turbulence inlet from the measured profile) and validated against those points, providing the full time-resolved fields. The dataset comprises 497 snapshots of pressure and the three velocity components, provided in .vtu (native solver mesh) and .npy (equispaced grid of 198 × 198 × 21 points at Δxy = 2 m, Δz = 5 m, with building mask) formats. - -[[Link to EchoNet-Dynamic](https://www.aij.or.jp/%20jpn/publish/cfdguide/index_e.htm)] - -## Urban LES database — AIJ Case K (regular 9×9 cubic array) -This database contains data from a three-dimensional large-eddy simulation of turbulent flow through a regular array of 81 cubic buildings in 9 rows × 9 columns, of side H = 0.06 m and pitch 2H, reproducing the AIJ Case K (ArraysC) wind tunnel benchmark at model scale (U_H = 2.33 m/s, Re ≈ 9800). Unlike the benchmark, which reports time-averaged velocity and turbulent kinetic energy at 203 probe points, the simulation was run with OpenFOAM (WALE subgrid-scale model, synthetic-turbulence inlet with the measured anisotropic Reynolds stresses), validated against those points, and rescaled by a factor of 400 (H = 24 m, pitch 48 m). The dataset comprises 299 snapshots of pressure and the three velocity components, provided in .vtu (native solver mesh) and .npy (equispaced grid of 205 × 205 × 21 points at Δxy = 2 m, Δz = 5 m, with building mask) formats. - -[[Link to EchoNet-Dynamic](https://www.aij.or.jp/%20jpn/publish/cfdguide/index_e.htm)] - -## Urban LES database — AIJ Case L (regular 14×9 cubic array) -This database contains data from a three-dimensional large-eddy simulation of turbulent flow through a regular array of 126 cubic buildings in 14 rows × 9 columns, of side H = 0.06 m and pitch 2H, reproducing the AIJ Case L (ArraysCT) wind tunnel benchmark at model scale (U_H = 1.34 m/s). Unlike the benchmark, which was conducted under weakly non-isothermal conditions (Rib ≈ −0.04) and reports time-averaged point measurements, the simulation was run isothermally with OpenFOAM (WALE subgrid-scale model, synthetic-turbulence inlet from the measured profile), validated against the benchmark velocities, and rescaled by a factor of 400 (H = 24 m, pitch 48 m). The dataset comprises 299 snapshots of pressure and the three velocity components, provided in .vtu (native solver mesh) and .npy (equispaced grid of 325 × 205 × 21 points at Δxy = 2 m, Δz = 5 m, with building mask) formats. +## Urban LES database — AIJ benchmark geometries +This database contains data from three-dimensional large-eddy simulations of turbulent flow over four urban configurations taken from the AIJ wind tunnel benchmark collection: two real urban layouts (a city district and a university campus, the latter including terrain elevation) and two regular arrays of cubic buildings. Specifically, it provides time-varying flow fields for the Niigata city district and for a real university campus, together with the canonical wind tunnel configurations of 9 × 9 and 14 × 9 equispaced cubic buildings. Unlike the benchmark, which reports time-averaged measurements at sparse probe points, the simulations were run with OpenFOAM using a WALE subgrid-scale model and a synthetic-turbulence inlet built from the measured wind tunnel profiles, and were validated against those measurements, so that the full time-resolved flow fields are made available. Each case provides a time series of snapshots of pressure and the three velocity components, in .vtu (native solver mesh) and .npy (interpolated onto an equispaced grid, at a common spatial resolution and accompanied by a static building occupancy mask) formats. -[[Link to EchoNet-Dynamic](https://www.aij.or.jp/%20jpn/publish/cfdguide/index_e.htm)] +[[Link to AIJ data sets](https://www.aij.or.jp/%20jpn/publish/cfdguide/index_e.htm)] -## Urban LES database — AIJ Case M (TPU university campus) -This database contains data from a three-dimensional large-eddy simulation of turbulent flow over a real university campus in Atsugi (Japan), reproducing the geometry of the AIJ Case M (TPU) wind tunnel benchmark at full scale (domain of about 1080 m × 720 m, tallest building 65 m, terrain elevation included). Unlike the benchmark, which reports time-averaged measurements at sparse points for several wind directions, the simulation was run for a single direction (180°, U_R = 4.08 m/s at H_R = 41 m) with OpenFOAM (WALE subgrid-scale model, synthetic-turbulence inlet from the measured profile) and validated against those measurements. The dataset comprises 450 snapshots of pressure and the three velocity components, provided in .vtu (native solver mesh) and .npy (equispaced grid of 329 × 292 × 21 points at Δxy = 2 m, Δz = 5 m, with building mask) formats. +## Urban wind tunnel database — EWTL Hamburg reference data sets +This collection gathers the reference data sets of the Environmental Wind Tunnel Laboratory of the University of Hamburg, comprising boundary layer wind tunnel measurements of flow and pollutant dispersion over idealized and realistic urban and industrial layouts, including the CEDVAL and CEDVAL-LES compilations, the COST Action ES1006 cases (Michelstadt and CUTE) and the COST Action 732 cases (MUST and Oklahoma City), with variations of wind direction, source location and release type. Unlike CFD databases, the data are physical measurements taken point by point with laser Doppler anemometry and fast flame ionization detectors, so the spatial coverage differs from case to case: some data sets sample a three-dimensional arrangement of measurement positions, whereas others are restricted to two-dimensional planes or to vertical profiles at selected locations, generally denser close to the ground. In all cases the flow is characterized in a statistically stationary sense, through time-averaged statistics (mean values, variances and higher moments, and in some cases point time series at individual probes). The data are distributed as tabulated measurement files and are password protected, available on request from the laboratory. -[[Link to EchoNet-Dynamic](https://www.aij.or.jp/%20jpn/publish/cfdguide/index_e.htm)] +[[Link to EWTL data sets](https://www.mi.uni-hamburg.de/en/arbeitsgruppen/windkanallabor/data-sets.html)]