souravsud/wind-cfd-trial

Wind Flow Over Complex Terrain Dataset

A large-scale dataset of steady-state RANS wind flow simulations over real-world complex terrain, designed for training machine learning models for wind resource assessment and atmospheric flow prediction.

Overview

Parameter	Value
Number of terrain locations	~1000
Wind directions per terrain	2 (random)
Total simulation cases	~10,000
Cropped grid per case	~298 × 298 × 64
Horizontal resolution (AOI)	~30 m
Vertical extent	~500 m AGL
Flow variables	U (3-component), p, k, ε
Surface variables	DEM, roughness (z₀), height AGL
Reference wind speed	10 m/s at 100 m height
Atmospheric stability	Neutral
Solver	OpenFOAM simpleFoam (RANS, k-ε)
Format	Zarr (xarray-compatible)

Quick Start

Load a single case

import xarray as xr

ds = xr.open_zarr("data/case_name.zarr")

# 3D wind field
Ux = ds['Ux'].values  # (ni, nj, nk) array, m/s
Uy = ds['Uy'].values
Uz = ds['Uz'].values

# Terrain
dem = ds['dem'].values        # (ni, nj), meters above sea level
z0 = ds['roughness'].values   # (ni, nj), aerodynamic roughness in meters

# Height above ground
h_agl = ds['h_agl'].values    # (ni, nj, nk), meters

# Metadata
print(ds.attrs['case_id'])
print(ds.attrs['rotation_deg'])   # wind direction
print(ds.attrs['converged'])      # simulation convergence flag

Load from Hugging Face directly

from huggingface_hub import hf_hub_download
import xarray as xr
import os

# Download a single case
local_path = hf_hub_download(
    repo_id="souravsud/wind-terrain-cfd",
    filename="data/case_name.zarr",
    repo_type="dataset",
    local_dir="./cache/"
)

ds = xr.open_zarr("./cache/data/case_name.zarr")

PyTorch DataLoader

See examples/02_dataloader_pytorch.py for a ready-to-use torch.utils.data.Dataset class.

Data Description

Per-case Zarr store contents

Variable	Shape	Units	Description
X	(ni, nj, nk)	m	UTM easting of cell centre
Y	(ni, nj, nk)	m	UTM northing of cell centre
Z	(ni, nj, nk)	m	Elevation above MSL
Ux	(ni, nj, nk)	m/s	Velocity x-component (UTM east)
Uy	(ni, nj, nk)	m/s	Velocity y-component (UTM north)
Uz	(ni, nj, nk)	m/s	Velocity z-component (vertical)
p	(ni, nj, nk)	m²/s²	Kinematic pressure (p/ρ)
k	(ni, nj, nk)	m²/s²	Turbulent kinetic energy
epsilon	(ni, nj, nk)	m²/s³	Turbulent dissipation rate
dem	(ni, nj)	m	Ground elevation (MSL)
roughness	(ni, nj)	m	Aerodynamic roughness length z₀
h_agl	(ni, nj, nk)	m	Height above ground level

Coordinate system

• X, Y: UTM coordinates. The EPSG code is stored in ds.attrs['utm_epsg'].
• Z: Absolute elevation above mean sea level (MSL), not height above ground.
• h_agl: Pre-computed height above ground: h_agl[i,j,k] = Z[i,j,k] - dem[i,j].
• The mesh is terrain-following (curvilinear). At each (i,j) column, Z increases with k but follows the terrain surface. Horizontal coordinates vary slightly with k.

Velocity scaling

All simulations use a reference velocity of 10 m/s at 100 m height under neutral atmospheric stability. Since the governing equations (incompressible RANS) are linear in velocity for neutral conditions, results can be scaled to any reference wind speed:

V_ref_desired = 8.0  # m/s
scale = V_ref_desired / 10.0

U_scaled = U_dataset * scale
p_scaled = p_dataset * scale**2
k_scaled = k_dataset * scale**2
epsilon_scaled = epsilon_dataset * scale**3

This is a feature, not a limitation — it means the dataset effectively covers all wind speeds.

Wind direction

Each case has a specific wind direction stored in ds.attrs['rotation_deg']. This is the angle (in degrees) by which the terrain was rotated to align the inlet boundary with the desired wind direction. The velocity components (Ux, Uy) are in the rotated UTM frame corresponding to that case.

Convergence quality

Each case includes convergence information:

• ds.attrs['converged']: Boolean flag (True if all residuals < 10⁻³)
• ds.attrs['residual_Ux'], ds.attrs['residual_p'], etc.: Final residual per field
• ds.attrs['iterations']: Number of solver iterations

The metadata/case_index.csv file contains convergence data for all cases, allowing easy filtering.

Dataset Structure

wind-terrain-cfd/
├── README.md                    # This file
├── data/
│   ├── case_0001.zarr/          # One Zarr store per simulation
│   ├── case_0002.zarr/
│   └── ...
├── metadata/
│   ├── case_index.csv           # Master index (lat, lon, wind_dir, converged, ...)
│   └── dataset_summary.json     # Aggregate statistics
└── examples/
    ├── 01_load_single_case.py
    ├── 02_dataloader_pytorch.py
    └── 03_velocity_scaling.py

Generation Pipeline

The dataset was generated using an automated pipeline:

1. Terrain fetching: terrain-fetcher — downloads DEM (Copernicus GLO-30) and land cover (ESA WorldCover) data
2. Mesh generation: terrain_following_mesh_generator — structured terrain-following mesh for OpenFOAM
3. Boundary conditions: ABL_BC_generator — neutral atmospheric boundary layer inlet profiles
4. Job management: taskManager — SLURM job submission and monitoring
5. Orchestration: CFD-dataset — end-to-end pipeline coordination

Citation

If you use this dataset in your research, please cite:

@dataset{sud2026windterrain,
  author = {Sud, Sourav},
  title = {Wind Flow Over Complex Terrain Dataset},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/souravsud/wind-terrain-cfd}
}

License

This dataset is released under the CC BY 4.0 license.