Aila AI model powered by LUMI advances the future of weather forecasting

AI-based weather models are transforming the way weather forecasts are produced. The Finnish Meteorological Institute’s Aila AI model aims to complement existing weather forecasting methods and improve regional forecasts, particularly in Finland and Northern Europe. Aila is being developed alongside physics-based numerical weather prediction models as a new forecasting tool.

Traditional weather forecasting models calculate atmospheric evolution using physical equations. In data-driven AI models, by contrast, a neural network is trained on historical weather situations and analysis datasets to generate forecasts quickly once the computationally intensive training phase is complete.

Aila was trained using the ERA5 reanalysis dataset produced by the European Centre for Medium-Range Weather Forecasts (ECMWF), along with analysis datasets from ECMWF’s Integrated Forecasting System (IFS) model and the high-resolution regional MEPS model. These datasets teach the model to recognize weather pattern evolution and generate forecasts tailored to conditions in Finland and Northern Europe.

Aila succeeded in forecasting the cold January

January 2026 provided a demanding test case for the model. Finland was colder than usual, and average temperatures in parts of Lapland fell below -20°C. Cold winter days and the associated atmospheric inversion conditions are challenging for traditional forecasting models.

Aila and other AI models tested, such as the Norwegian Meteorological Institute’s Bris model, performed extremely well at forecasting cold temperatures in Finland. When compared with observed station temperatures, Aila’s forecasts provided added value, particularly for forecast lead times of less than two days, relative to numerical weather prediction models. Further development is still needed in areas such as uncertainty estimation, the coldest extreme events, regionally variable phenomena such as precipitation, coastal areas, and longer-range forecasts spanning three to ten days.

The aim is for AI models to provide additional support for forecasting work in the future, especially during regionally complex and rapidly changing weather situations.

Open development supports international collaboration

Aila has been in testing use at the Finnish Meteorological Institute for approximately one year, and its development is now being opened more broadly to the international weather and AI communities. The model weights, which represent the computational parameters the AI model learned during training, are now openly available on the Hugging Face platform. This open release facilitates comparison, testing, and further development of the model and supports more open collaboration in AI-based weather forecasting.

Aila’s technical implementation is based on the Anemoi software framework, developed by ECMWF together with national meteorological services. Anemoi provides an open and modular foundation for building, training, and operationally deploying machine learning-based weather forecasting systems.

The Aila model uses a graph neural network and a stretched-grid approach. The stretched grid makes it possible to run a global weather model while providing greater forecast accuracy over Northern Europe than elsewhere.

Training Aila requires substantial GPU computing resources, meaning high-performance computing facilities suitable for AI workloads. For this purpose, the project has utilized the GPU capacity of the LUMI supercomputer, which researchers from the Finnish Meteorological Institute have accessed through Finland’s national LUMI Extreme Scale allocation. These computing resources make it possible to train neural network models of Aila’s scale and complexity.

Aila brings together research, operations, and the European community

The Finnish Meteorological Institute’s AI-based weather forecasting team currently consists of around ten experts. The work combines the institute’s research expertise with its operational forecasting capabilities. Internationally, the development effort is linked to ECMWF’s Machine Learning Project and to Nordic collaboration initiatives, including cooperation with the Norwegian Meteorological Institute’s Bris model, whose model architecture serves as the foundation for Aila.

Aila is part of a broader transformation in which AI, large observational and analytical datasets, and powerful computing resources are creating new opportunities for environmental modelling and forecasting. The Finnish Meteorological Institute participates in the Academy of Finland’s FAME Flagship program, and from that perspective Aila combines advanced AI research, high-performance computing, international collaboration, and the operational development of forecasting services.

In 2025, the Anemoi community received the European Meteorological Society (EMS) Technology Achievement Award. The award highlighted Anemoi’s scientific and technological significance and demonstrated how open European collaboration can accelerate the adoption of next-generation weather forecasting methods. The Finnish Meteorological Institute’s Aila model is a concrete example of how joint development efforts are moving from research into practical forecasting applications.

The original version of this text was published in Finnish on the Finnish Meteorological Institute’s website. 

Image: Adobe Stock