Prithvi WxC
| Property | Value |
|---|---|
| Parameters | 2.3 Billion |
| License | CDLA-Permissive-2.0 |
| Paper | arXiv:2409.13598 |
| Framework | PyTorch |
What is prithvi.wxc.2300m.v1?
Prithvi WxC is a state-of-the-art foundation model for weather and climate applications, trained on 160 different variables from NASA's MERRA-2 dataset. This particular version features a 50% masking ratio and is designed for both atmospheric state reconstruction and future forecasting capabilities.
Implementation Details
The model operates by processing data from two timestamps as input to generate a single output timestamp, which can be either a reconstruction or a future prediction. During pretraining, the model was configured with input time deltas ranging from -3 to -12 hours, and forecast lead times varying from 0 to 24 hours.
- Flexible time delta between input timestamps (-3, -6, -9, -12 hours)
- Variable forecast lead times (0, 6, 12, 24 hours)
- 50% masking ratio for robust feature learning
- Built on PyTorch framework
Core Capabilities
- Atmospheric state reconstruction from partial information
- Weather forecasting across multiple timescales
- Zero-shot reconstruction capabilities
- Supports downstream tasks like downscaling and gravity wave analysis
Frequently Asked Questions
Q: What makes this model unique?
Prithvi WxC stands out for its dual capability in both reconstruction and forecasting tasks, processing 160 different atmospheric variables. Its architecture allows for flexible time deltas and forecast horizons, making it versatile for various weather and climate applications.
Q: What are the recommended use cases?
The model is recommended for generic use cases that don't exclusively focus on forecasting. It's particularly well-suited for atmospheric state reconstruction, downscaling applications, and gravity wave analysis. The model has been successfully applied in projects like the IBM-Granite geospatial downscaling and NASA's gravity wave studies.
