Artificial Intelligence for Numerical Weather Prediction

In 2023 the Met Office and The Alan Turing Institute launched the Artificial Intelligence for Numerical Weather Prediction (AI4NWP) Partnership - a strategic partnership to explore how artificial intelligence can help shape the future of weather forecasting. Bringing together world-leading expertise in meteorology, weather prediction and data science, the collaboration is helping us understand how machine learning can work alongside physics-based modelling and expert meteorological judgement to deliver future operational forecasting capability. 

Through the AI4NWP Partnership, researchers are developing new approaches to weather prediction, building scientific evidence and operational understanding needed to support the responsible, transparent and safe use of AI in weather and climate services, helping build public confidence in these emerging technologies.  

FastNet - An experimental global Artificial Intelligence-based weather forecast system

One of the partnership's most significant achievements to date, and its current focus, is the co-development of FastNet: a Machine Learning Weather Prediction (MLWP) model. FastNet provides a practical testbed for exploring how advances in artificial intelligence can contribute to future forecasting systems. 

The name FastNet will be familiar to many as one of the 31 sea areas covered by the Shipping Forecast. It's also a nod to the Met Office’s founder, Vice-Admiral Robert FitzRoy, ounder of the Met Office, whose work helped lay the foundations for marine forecasting and the protection of life and property at sea. 

Since the partnership began, the collaboration has progressed from early research and prototype models to peer-reviewed scientific publications and pre-operational evaluation. Recent research has demonstrated how meteorological expertise can be incorporated into the training of AI weather prediction systems, helping improve the physical realism of forecasts while maintaining strong forecast performance. 

This work is helping us better understand how AI can be applied safely, responsibly and transparently in weather forecasting, while building confidence in the evidence needed to support future operational use. 

Developing the next generation of weather forecasting 

As the UK's national meteorological service, the Met Office continually invests in the science and technology needed to improve weather forecasting and climate prediction. Machine learning presents an exciting opportunity to complement existing forecasting approaches by producing useful forecasts significantly faster and at lower computational cost than traditional methods. 

The goal is not to completely replace physics-based weather prediction. Physics-based models remain fundamental to our forecasting capability and provide the scientific foundation upon which machine learning systems are developed and evaluated. Instead, our research is focused on understanding how the strengths of machine learning, physics-based modelling and expert meteorological judgement can be combined to deliver better outcomes. 

This work is helping us explore what the future forecasting landscape could look like, where AI enhances capability while maintaining the scientific rigour, reliability and trust that people expect from the Met Office. 

 From research partnership to real-world capability 

Since its launch, the Artificial Intelligence for Numerical Weather Prediction project has delivered significant scientific progress. The Met Office and Alan Turing Institute partnership has developed FastNet, published peer-reviewed research, and begun evaluating machine learning forecasts alongside existing forecasting systems. Research has shown how incorporating physical principles into the training process can help machine learning forecasts better represent important meteorological features, including storm structures and sharp weather fronts. These advances are helping improve physical consistency and build confidence in future AI forecasting systems. 

Whilst this partnership continues to develop FastNet, with current developments looking to produce both global and regional UK forecasts, the Met Office are also working towards integrating the model into our operational workflows. This includes both real-time and retrospective assessment and evaluation of AI-based forecasts that will allow us to assess how best these can contribute to future forecasting capability. 

Future research will continue to explore how machine learning, physics-based modelling and meteorological expertise can work together to improve weather forecasting. By combining scientific excellence, operational experience and responsible innovation, we are helping shape the future of weather prediction for the benefit of society. 

Trust, transparency and responsible AI 

Weather forecasts support decisions that affect lives, livelihoods and infrastructure. For that reason, any future use of artificial intelligence within forecasting systems must be guided by the same scientific standards, evaluation processes and commitment to public service that underpin all Met Office forecasts and warnings. 

Our research is focused not only on improving forecast performance, but also on understanding how AI systems can be developed responsibly and transparently. This includes ensuring forecasts remain physically realistic, rigorously evaluating model performance, understanding limitations, and maintaining appropriate human oversight. 

Expert meteorologists and scientists remain central to this work. Their knowledge plays a critical role in evaluating new forecasting approaches, interpreting results and ensuring that scientific evidence guides future operational decisions. 

As research continues, our focus remains on building trust alongside capability, ensuring any future use of AI strengthens confidence in the forecasts and warnings people rely upon every day.

Experimental AI-based FastNet weather forecast system in near real-time

FastNet web disclaimer:The Met Office is running the experimental Artificial Intelligence-based FastNet weather forecast system in near real-time.  An illustration of this output is published daily, however, as this is a non-operational trial, intended for research purposes only – provided simply by way of demonstration and for comparison against existing products – this is not to be used commercially or as a decision-making aid. (Images may, at times, also be unavailable).