| Property | Value |
|---|---|
| Model Type | Encoder-processor-decoder model |
| Developer | ECMWF |
| License | CC BY 4.0 |
| Checkpoint Size | 1.19 GB |
| Paper | arXiv:2406.01465 |
What is aifs-single-0.2.1?
AIFS (Artificial Intelligence Forecasting System) is a sophisticated weather forecasting model developed by ECMWF that combines graph neural networks with transformer architecture. It's designed to produce highly skilled forecasts for upper-air variables, surface weather parameters, and tropical cyclone tracks, running four times daily alongside ECMWF's traditional numerical weather prediction models.
Implementation Details
The model utilizes a GNN encoder-decoder architecture with a sliding window transformer processor. It's trained on ERA5 re-analysis data and ECMWF's operational analyses, implementing three training phases: pre-training on ERA5 (1979-2020), fine-tuning with rollout training, and additional fine-tuning on operational IFS NWP analyses. The model employs AdamW optimizer with mixed precision training across 64 A100 GPUs.
- Trains on multiple atmospheric variables across 13 pressure levels
- Implements area-weighted MSE loss function with variable-specific scaling
- Supports parallel processing for high-resolution data handling
- Uses flash_attention for improved computational efficiency
Core Capabilities
- 6-hour forecast generation with auto-regressive extension to 72 hours
- Comprehensive atmospheric state prediction including temperature, pressure, wind components
- Surface weather parameter forecasting including precipitation and temperature
- Integration with ECMWF's operational forecasting system
Frequently Asked Questions
Q: What makes this model unique?
AIFS combines modern deep learning architectures with traditional meteorological expertise, offering competitive forecast skill compared to physics-based models while being computationally efficient. It's one of the few AI weather models running operationally at a major forecasting center.
Q: What are the recommended use cases?
The model is particularly suited for medium-range weather forecasting, including surface weather parameters, upper-air variables, and tropical cyclone tracking. It's designed for operational meteorological applications and research purposes, with results available through ECMWF's open data policy.