What is AIFS

The Artificial Intelligence Forecasting System (AIFS) is ECMWFs interpretation of a data-driven weather forecasting system.

The model may change, but currently we're working with graph neural networks (GNNs) which starts with the representation of atmospheric data as a graph.

In this context, each geographical location, such as grid cells (or weather stations later), becomes a node in the graph. The edges between these nodes represent the spatial relationships and interactions between different locations. For example, if two locations influence each other's weather patterns, there is an edge connecting them.

The key elements of our graph representation include:

  • Input Nodes: Each node corresponds to a specific geographical location and holds information about the weather variables at that location. Which is then passed to other nodes in the network.

  • Hidden Nodes: The network containes a processing grid where the nodes do not represent the data locations but rather present the information connection on different levels of coarseness.
  • Output Nodes: The output nodes map the information back to the data domain.
  • Edges: Edges capture the connections between nodes, indicating the influence one location has on another. That's why our graph has both short as well as tele-connections.

  • Trainable Edges: The edge features contain information about the local weather conditions at each edge.

AI-Models for Inference

We have the open-source tool ai-models for inference of our models with an aifs plugin.

AI-Models

PrepML

We have our internal tool for validation and verification of data-driven forecasts:

prepml




Getting Started

Here's a quick start guide for AIFS development.

Code

The code repository lives here:

https://github.com/ecmwf-lab/aifs-mono

And we have an installation guide here:

How to install AIFS

We use pre-commit hooks to ensure code quality which is mentioned there as well!

Configs

We use configs to define default values for our model:

Using AIFS / Hydra Configuration