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

Area of Science:

• Atmospheric Science
• Computational Fluid Dynamics
• Artificial Intelligence

Background:

• Accurate near-surface wind field prediction is crucial for many applications.
• Current models average out terrain-induced wind features due to coarse resolution.
• There is a need for high-resolution wind field models that incorporate terrain effects.

Purpose of the Study:

• To introduce FuXi-CFD, a machine learning framework for generating high-resolution 3D near-surface wind fields.
• To demonstrate the framework's ability to infer full 3D wind fields from limited inputs.
• To validate the model's performance against CFD simulations and real-world observations.

Main Methods:

• Developed FuXi-CFD, a machine learning framework utilizing coarse atmospheric inputs and high-resolution terrain data.
• Trained the model on a large dataset generated using computational fluid dynamics (CFD) across diverse terrains.
• The model infers 3D wind fields, including vertical velocity and turbulence, from horizontal wind inputs.

Main Results:

• FuXi-CFD generates 3D wind fields at 30-meter horizontal resolution.
• The framework achieves accuracy comparable to CFD simulations.
• Inference time is reduced from hours to seconds, enabling real-time applications.
• The model demonstrates strong generalization to real-world conditions, validated by wind-tower observations.

Conclusions:

• FuXi-CFD provides a significant advancement in fine-scale wind field prediction.
• The framework enables real-time, high-resolution wind field reconstruction for terrain-sensitive applications.
• Applications include wind turbine siting, power forecasting, and wildfire spread modeling.