FastNet

Machine learning weather prediction for the future of forecasting

FastNet is a machine learning weather prediction model being developed by the Met Office and The Alan Turing Institute. It is helping us understand how artificial intelligence can support the future of weather forecasting while maintaining the scientific rigour, reliability and trust that people expect from Met Office forecasts and warnings.  

Named after the Fastnet sea area featured in the Shipping Forecast, FastNet combines advances in machine learning with meteorological science to explore new ways of predicting the weather.  

What is FastNet? 

FastNet is a machine learning weather prediction model designed to produce weather forecasts by learning patterns from historical atmospheric data. Unlike physics-based weather models, which simulate the atmosphere using mathematical representations of physical processes, FastNet learns from decades of weather observations, models and analyses to generate forecasts.  

The model is being developed to produce global forecasts up to two weeks ahead and regional forecasts for the UK and other areas of interest.  

FastNet forms part of the Met Office's wider work to explore how machine learning can contribute to future forecasting capability alongside existing modelling systems and expert meteorological judgement.  

Why are we developing FastNet? 

Machine learning weather prediction has the potential to produce useful forecasts significantly faster and using less computing power than existing forecasting approaches. This could create new opportunities to improve forecasting capability, increase forecast availability and support better decision-making.  

As severe weather becomes more frequent in a changing climate, faster and more useful forecasts could help people, businesses and emergency responders make better decisions to protect lives, livelihoods and infrastructure.  

FastNet is also helping the Met Office develop the expertise, evidence and operational understanding needed to assess how machine learning weather prediction could be used responsibly in future forecasting systems.  

How does FastNet work? 

FastNet uses a type of artificial intelligence known as a graph neural network. It is trained using large volumes of historical atmospheric data and learns how weather patterns evolve over time.  

The model takes an atmospheric state as input and predicts how conditions will change over the coming hours. These predictions are then used to generate forecasts extending from days to weeks ahead.  

Researchers continue to develop the model to improve forecast skill, increase resolution and better represent important weather features such as fronts, storms and high-impact weather events.  

The FastNet journey 

FastNet's development has progressed through several stages.  

FastNet v1 was developed to demonstrate that the Met Office and The Alan Turing Institute could successfully build and train a global machine learning weather prediction model. It provided the foundation for future development and helped build capability. skills and experience within the programme and the team.  

FastNet v1.1 introduced new approaches designed to improve the physical realism of forecasts. Research showed that incorporating physical principles into model training could improve the representation of weather systems while maintaining strong forecast performance.  

FastNet v2, currently in development, is a family of models designed to produce both global and regional UK forecasts. It represents the next major step towards understanding how machine learning weather prediction could contribute to future operational forecasting capability. This and subsequent versions of the model are be built in the Anemoi framework, developed by a collaboration of European Meteorological Organisations. The use of Anemoi allows the Met Office and the Alan Turing Institute to work as part of a growing modelling community, whilst delivering the UK a sovereign operational capability. 

Evaluating FastNet in real-world forecasting 

FastNet is currently being evaluated alongside existing forecasting systems to better understand its strengths, limitations and potential future applications. Researchers and operational meteorologists routinely assess its performance against established forecasting approaches, helping build confidence in how machine learning weather prediction could contribute to future forecasting capability. 

This evaluation goes beyond forecast accuracy alone. Researchers are also examining how well FastNet represents important weather features, including storms, weather fronts and other high-impact events, as well as assessing the physical realism and consistency of its forecasts. Recent research has shown that incorporating meteorological expertise directly into the training process can improve the physical consistency of forecasts while maintaining strong predictive performance. 

Operational meteorologists play an important role in this process. Their expertise helps evaluate how machine learning forecasts perform in real-world forecasting situations, identify strengths and limitations, and ensure new approaches are assessed against the high standards required for operational weather forecasting. 

Before operational use, FastNet will continue to undergo scientific evaluation, operational assessment and assurance to ensure it is reliable, safe and supports and maintains confidence in Met Office forecasts and warnings.