FuXi-CFD Dataset

Overview

This dataset accompanies the paper:

Reconstructing fine-scale 3D wind fields with terrain-informed machine learning

It contains large-scale CFD-generated wind field simulations used for training and evaluating the FuXi-CFD model described in the paper.

The dataset includes over 12,000 cases.

---

Dataset Organization

The dataset is distributed as compressed archive files:

FuXi-CFD-dataset-part01.tar
FuXi-CFD-dataset-part02.tar
...
FuXi-CFD-dataset-part13.tar

Each case follows the structure:

case_000001/
    inputs.npz
    outputs.npz

---

Inputs (inputs.npz)

The inputs.npz file contains four variables:

dem

• Shape: (300, 300)
• Unit: meters (m)
• Horizontal resolution: 30 m

Digital Elevation Model (DEM) representing terrain elevation.

---

roughness

• Shape: (300, 300)
• Unit: meters (m)
• Horizontal resolution: 30 m

Surface roughness length (z₀) map representing aerodynamic terrain roughness.

---

u_100m

• Shape: (9, 9)
• Unit: m s⁻¹
• Horizontal resolution: approximately 1 km

Coarse-resolution zonal wind component at 100 m height.

---

v_100m

• Shape: (9, 9)
• Unit: m s⁻¹
• Horizontal resolution: approximately 1 km

Coarse-resolution meridional wind component at 100 m height.

---

Outputs (outputs.npz)

The outputs.npz file contains four high-resolution 3D variables.

All output variables have shape:

(27, 300, 300)

where:

• 27 corresponds to the number of vertical levels (in meters above ground level).
• 300 × 300 corresponds to 30 m horizontal resolution

The 27 vertical levels are:

5, 10, 15, 20, 25, 30, 35, 40, 45, 50, 55, 60, 65, 70, 75, 80, 85, 90, 95, 100,
106.5, 114.95, 125.94, 140.22, 158.78, 182.91, 214.29

---

u

• Unit: m s⁻¹

Zonal wind component (east–west).

---

v

• Unit: m s⁻¹

Meridional wind component (north–south).

---

w

• Unit: m s⁻¹

Vertical wind component.

---

k

• Unit: m² s⁻²

Turbulent kinetic energy (TKE).

---

Resolution Summary

---

Data Split

The dataset is provided as a single collection without predefined splits.

The official train/validation/test split used in the paper can be found in the associated publication.

Users are free to define custom splits for their own applications.

---

License

CC BY-NC 4.0

---

Citation

If you use this dataset, please cite:

Lin, C., et al. Reconstructing fine-scale 3D wind fields with terrain-informed machine learning, Nature Communications (2026).