Artificial Intelligence Weather Model AIFS

ECMWF releases its machine learning model as open-data

The open-data weather forecasting landscape is experiencing a breath of fresh air with an exciting development the European Centre for Medium-Range Weather Forecasts (ECMWF).

ECMWF Opens the Doors to AIFS: In a significant move towards open-data, ECMWF has made data from its cutting-edge artificial intelligence weather model (AIFS) publicly available.

Unlike traditional numerical models, AIFS leverages Graph Neural Networks (GNNs), similar to those used in AI image generation, but with significantly more data. This shift, made possible by recent advancements in computing power, allows AIFS to learn complex weather patterns with high accuracy.

In recent years, several AI weather models have emerged, with Google's GraphCast gaining attention for surpassing even the world-renowned IFS model. However, AIFS now takes the crown, demonstrating superior accuracy as shown in the provided forecast performance chart below. While all models perform well in the short term (first few days), AI models like AIFS excel in longer-range forecasts, exceeding 5 days.

This development is a significant step forward for AI in weather forecasting, paving the way for more accurate and accessible weather information for everyone.

Anomaly Correlation for 500 hPa Geopotential in the Northern Hemisphere for 2022. Source: ECMWF

While AI models like AIFS hold great promise, they currently have limitations. They provide a smaller range of weather variables and offer forecasts only in 6-hour intervals. This means they might not be ideal for specific applications demanding detailed information, such as short-term forecasting for agriculture or energy needs. In such cases, high-resolution local models like the European 3 km DWD ICON-D2 or the North American HRRR remain more suitable options.

Despite these limitations, the open availability of AIFS data is a significant step forward. The raw data is accessible on the ECMWF open-data server and also integrated into the Open-Meteo API. This allows users to easily compare AIFS forecasts with traditional models like the ECMWF IFS side-by-side, facilitating independent assessment of their relative accuracy.

Temperature Forecast for Berlin from IFS and AIFS in 0.25° resolution

While definitive pronouncements on which model is "more correct" are challenging, it seems AIFS aligns well with established models (chart below), offering a promising new perspective for weather forecasting.

Comparison with other Weather Models on Open-Meteo

Make sure to subscribe to get the latest news around weather forecasting!