FuXi-CFD Model
Overview
This repository accompanies the paper:
Reconstructing fine-scale 3D wind fields with terrain-informed machine learning
It provides the pre-trained FuXi-CFD model used in the study, exported in ONNX format, together with a complete inference example.
Version: v1.0
Framework: ONNX (runtime inference)
Model Description
FuXi-CFD is a terrain-informed deep learning model designed to reconstruct fine-scale three-dimensional wind fields from coarse atmospheric inputs and high-resolution terrain information.
Model Inputs
The model expects an inputs.npz file containing:
dem — terrain elevation (300 × 300, float32, meters)roughness — surface roughness length (300 × 300, float32, meters)u_100m — coarse zonal wind at 100 m (9 × 9, float32, m s⁻¹)v_100m — coarse meridional wind at 100 m (9 × 9, float32, m s⁻¹)
All variables must be provided in physical units (no normalization applied by the user).
Normalization parameters used during training are included in normalization/.
Model Outputs
The model produces a file prediction.npz containing:
u — zonal wind component (27, 300, 300), m s⁻¹v — meridional wind component (27, 300, 300), m s⁻¹w — vertical wind component (27, 300, 300), m s⁻¹k — turbulent kinetic energy (27, 300, 300), m² s⁻²
The vertical levels correspond to the 27 non-uniform heights described in the associated dataset documentation.
Quick Start (Inference Example)
1cd inference_example
2python scripts/infer.py \
3 --model ../model/fuxicfd_model.onnx \
4 --input data/inputs.npz \
5 --output data/prediction.npz
Outputs are saved as data/prediction.npz with keys: u, v, w, k.
Repository Structure
model/ — exported ONNX weightsinference_example/ — complete preprocessing → inference → postprocessing pipeline
normalization/ — training-time normalization parametersscripts/ — runnable inference scriptsutils/ — helper functions
License
CC BY-NC 4.0
Citation
If you use this model, please cite:
Lin, C., et al. Reconstructing fine-scale 3D wind fields with terrain-informed machine learning, Nature Communications (2026).
