WindINR: Neural Framework for Fast Local Wind Queries in Complex Terrain
WindINR is a framework designed for understanding high-resolution local wind conditions, especially in complex landscapes, by using limited data points. It features a latent-conditioned decoder that transforms basic terrain information, a low-resolution background, and specific query coordinates into detailed wind data. This system effectively separates general representation learning from adjustments tailored to specific samples, allowing for quick corrections during use. During training, a specialized encoder creates a reference state with high-resolution data, while a practical predictor establishes an initial state using only the inputs available at the time. The differences between these states inform a data-driven correction. This approach enables rapid wind estimations for certain locations and altitudes for a given forecast time, avoiding the need for dense grid forecasts. You can find this research on arXiv, ID 2605.09511.
Key facts
- WindINR is a latent-state implicit neural representation framework.
- It enables continuous high-resolution local wind query and sparse-observation correction.
- Maps static terrain descriptors, low-resolution background field, and continuous query coordinates to high-resolution wind state.
- Uses a latent-conditioned decoder.
- Separates reusable representation learning from sample-specific latent-state correction.
- Privileged encoder infers reference latent state from high-resolution supervision.
- Deployable latent predictor estimates initial latent state from inference-time inputs.
- Published on arXiv with ID 2605.09511.
Entities
Institutions
- arXiv