Enterprise AI Analysis

FuXi-CFD: Bridging Physics and AI

The FuXi-CFD framework integrates large-scale computational fluid dynamics (CFD) data with deep learning to create a terrain-response operator. A massive dataset of over 12,000 high-resolution CFD simulations, covering diverse terrain and inflow conditions in southeastern China, was generated to train the model. This dataset ensures a physically consistent foundation for learning terrain-induced wind variability.

The deep learning model takes as input 100m wind components (u, v) at 1km resolution and high-resolution terrain elevation and roughness maps at 30m resolution. Utilizing a Vision Transformer backbone, it captures complex spatial correlations and long-range dependencies. The model then reconstructs full 3D wind fields (u, v, w components) and turbulent kinetic energy (k) across 27 vertical levels, up to 214m. A hybrid loss function combining spatial and frequency-domain terms ensures both local accuracy and large-scale structural coherence. This approach allows the model to infer full 3D wind structures from limited coarse inputs, making it applicable where traditional CFD would be infeasible.

Key Findings & Performance

FuXi-CFD demonstrates high predictive skill, reproducing horizontal wind components with high fidelity and minimal scatter. While vertical velocity shows slightly larger spread, it remains well-constrained, effectively recovering terrain-induced updraft and downdraft structures. Turbulent kinetic energy, inferred indirectly, shows the largest dispersion as expected, given its sensitivity to small-scale shear.

Performance consistently improves with height, with the 100m level showing the strongest agreement, as upper-level winds are more anchored by coarse-scale inputs. The model robustly captures both large-scale momentum distribution and fine-scale terrain-modulated responses across various heights and terrains. Crucially, FuXi-CFD achieves CFD-comparable accuracy while reducing inference time from hours to seconds, enabling real-time applications. Validation against independent tall-tower observations in Europe confirmed strong cross-regional generalization, with significant error reductions compared to ERA5-based baselines, particularly in high-wind conditions and pronounced terrain effects.

Real-world Applications

The capability of FuXi-CFD to reconstruct fine-scale 3D wind fields in real-time offers transformative potential for several high-impact applications:

• Wind Turbine Siting & Power Forecasting: Accurate 30m resolution wind data is crucial for optimizing wind farm layouts, predicting energy output, and managing grid stability, especially in complex topographies where wind behavior is highly localized.
• Wildfire Spread Modeling: Detailed 3D wind fields, including vertical velocity and turbulence, provide critical inputs for more accurate wildfire behavior predictions, enabling better resource allocation and emergency response.
• Urban Risk Assessment & Air Quality: Fine-scale wind information is essential for understanding pollutant dispersion in urban environments and assessing wind loads on structures.
• Extreme Wind Assessment: Identifying localized extreme wind events, often terrain-induced, is vital for infrastructure design and safety, which is poorly resolved by coarse models.