The Artificial Intelligence Forecasting System (AIFS) is ECMWF's data-driven forecasting system.
Instead of solving physical equations of the atmosphere like the IFS, the AIFS uses machine learning and historical weather data to learn statistical representations of the atmosphere.
Architecture
The AIFS architecture consists of:
- Attention-based graph neural networks, which form the encoder and decoder.
- A sliding-window transformer, which acts as the processor.
These components work as follows:
- The encoder creates a simplified, compressed representation of the atmosphere from input fields.
- The processor learns how the atmosphere evolves over time.
- The decoder converts the updated atmospheric representation back to a high-resolution output grid.
Training
The AIFS has been trained on enormous volumes of historical weather data - the ERA5 reanalyses - to learn statistical patterns of atmospheric behaviour.
- AIFS Single was trained using ERA5 data for the years 1979 to 2022
- AIFS ENS was trained using ERA5 data from the years 1979 to 2017
The AIFS was then fine-tuned using ECMWF's operational analysis data from the years 2016-2023.
Inference (forecasting)
The AIFS runs alongside ECMWF's traditional physics-based Integrated Forecasting System (IFS).
Inference is the process by which the AIFS generates a weather forecast.
- Analysis fields from the IFS are used as input (initial conditions) for the AIFS.
- The analysis fields are re-gridded to N320 resolution for compatibility with the AIFS.
- Inference is conducted very rapidly (in minutes) on graphics processing units (GPUs).
- AIFS-Single generates 10-day forecasts four times per day.
- AIFS-ENS generates 15-day forecasts four times per day.
Further reading
Technical details about the model architecture, data and training process are provided in the arXiv pre-prints: