AIFS Single v1.0
| Property | Value |
|---|---|
| Model Type | Encoder-processor-decoder with GNN |
| Resolution | ~36km horizontal, 13 vertical levels |
| License | CC BY 4.0 |
| Paper | arXiv:2406.01465 |
| Checkpoint Size | 1.19 GB |
What is aifs-single-1.0?
AIFS Single v1.0 is the European Centre for Medium-Range Weather Forecasts' (ECMWF) first operationally supported AI-based weather forecasting system. It represents a significant advancement in applying machine learning to numerical weather prediction, utilizing a graph neural network encoder-decoder architecture combined with a sliding window transformer processor.
Implementation Details
The model is trained on ERA5 reanalysis and ECMWF operational analyses, providing forecasts four times daily. It operates at approximately 36km horizontal resolution with 13 pressure levels, making it highly efficient while maintaining impressive forecast accuracy.
- Trained using AdamW optimizer with mixed precision on 64 A100 GPUs
- Uses Flash Attention for improved computational efficiency
- Incorporates area-weighted mean squared error loss function
- Processes both atmospheric and surface parameters
Core Capabilities
- Produces skilled forecasts for upper-air variables and surface weather parameters
- Generates accurate tropical cyclone track predictions
- Provides new outputs including 100-meter winds and solar radiation
- Handles multiple atmospheric variables across different pressure levels
- Computes surface parameters including precipitation and cloud cover
Frequently Asked Questions
Q: What makes this model unique?
AIFS Single v1.0 is unique in being one of the first operational AI weather forecasting systems from a major meteorological center, combining traditional meteorological expertise with modern deep learning approaches. It's designed to run alongside conventional numerical weather prediction models, providing complementary forecasting capabilities.
Q: What are the recommended use cases?
The model is best suited for medium-range weather forecasting, particularly for predicting standard meteorological parameters like temperature, wind, and precipitation. It's especially valuable for operational meteorology, research purposes, and as a complementary tool to existing numerical weather prediction systems.
