How The Met Office is Using AI to Forecast the Future

AI is reshaping how we understand weather, climate and risk. For the Met Office, the UK’s national meteorological service, that shift is as much about sustainability as it is about accuracy.

Kirstine Dale, the Met Office's Chief AI Officer and Principal Fellow for Data Science, has spent the past two years embedding AI across the organisation. She describes this as “a journey that has been informative, exciting and challenging,” one that has overturned assumptions about public sector innovation.

“I have seen the public sector be agile and innovative in a way I really didn’t know was possible,” Kirstine says. That agility matters because weather and climate information underpins everything from aviation safety to energy planning and climate adaptation. The Met Office already runs one of the world’s most advanced numerical weather prediction systems, powered by a £1.2bn (US$1.6bn) supercomputing partnership with Microsoft.

Modelling the atmosphere

Traditional forecasts rely on a physics-based model of the atmosphere, using the laws of motion, mass and energy to simulate how weather evolves. Every day, around 215 billion observations from satellites, land stations and ocean sensors flow into this system, where they are assimilated into high‑resolution simulations.

Operational meteorologists then evaluate and interpret those outputs for decision‑makers, ranging from airport operators to local authorities managing flood risk. Kirstine notes that “92.5% of our next‑day temperature forecasts are accurate within two degrees,” a level of performance that has been steadily improving. Yet incremental gains are no longer enough, especially as climate change drives more frequent extremes and more complex risks. 

“We are constantly looking for ways to improve,” Kirstine explains, “and that’s why we have been keeping a very close eye on AI.”

In 2022, the Met Office created a data science framework to guide how AI and machine learning would be integrated into forecasting and services. At that stage, AI was expected to be an evolution, not a revolution, with careful experimentation and gradual deployment into parts of the value chain. Then, over the 2022–23 winter, everything changed.

AI models challenge the old paradigm

While Kirstine was planning a quiet Christmas exploring early large language models such as ChatGPT, the research landscape for AI weather prediction moved dramatically. Several powerful machine learning weather models, developed largely by major technology companies, were released and benchmarked against traditional systems. 

“These models showed a step change in performance,” Kirstine reveals. “They were genuinely competitive with physics‑based approaches.”

That raised a fundamental question for policymakers and the public – if AI can predict weather as well as conventional systems, do we still need billion‑pound supercomputers and specialist meteorologists? For a public body accountable for safety‑critical information, the answer required rapid, evidence‑based action.

The Met Office responded by launching its AI for Numerical Weather Prediction programme, pulling together related work into a coherent research and development effort. 

“We realised the race had started and we had ground to make up,” Kirstine continues. “We knew we couldn’t do it alone.”

FastNet: AI at the heart of forecasting

In October 2023, the Met Office formed a strategic partnership with The Alan Turing Institute, the UK’s national centre for data science and AI, to co‑develop a new AI model called FastNet.

FastNet is an experimental global weather prediction system based on machine learning, using a graph neural network architecture to model atmospheric patterns. Unlike physics‑based numerical models, which simulate the atmosphere from first principles, FastNet learns relationships directly from large volumes of historical and simulated data.

“The FastNet model is an experimental global AI‑based weather forecast,” Kirstine explains. “It’s completely different to the physics‑based models we’ve used traditionally.”

Early results show FastNet is now competitive with the leading machine learning weather models globally, despite starting from scratch just a few years ago. On internal benchmarking charts, FastNet now sits firmly in the leading pack, a visible symbol of how quickly public sector science can adapt. 

“This shows just how far we’ve come and how agile and innovative the public sector has had to be,” Kirstine says. 

Crucially, AI models such as FastNet are not only accurate but computationally far lighter than full‑scale physics simulations.

Speed, emissions and sustainable compute

Traditional numerical weather prediction requires vast supercomputing resources, consuming large amounts of energy to run models at high resolution and high frequency. By contrast, once a machine learning model is trained, generating new forecasts – known as inference – can be tens of thousands of times faster and much less energy‑intensive.

“AI models are not just slightly faster,” Kirstine stresses. “They’re tens of thousands of times faster than physics‑based models.”

Fast inference opens up three major sustainability advantages:

• First, it enables forecasts to be run closer to real time, so more of the latest observational data can be assimilated, improving relevance for climate‑sensitive sectors such as renewable energy and agriculture.
• Second, the lower computational load allows models to run on edge devices closer to users, rather than exclusively on centralised supercomputers. This can reduce data transfer, cut latency and open up resilient local forecasting for regions with weaker infrastructure.
• Third, and most directly, a lighter‑weight model consumes less power per forecast, reducing the operational carbon footprint of weather services. “The third advantage of this light computational load is that it’s more environmentally friendly,” Kirstine notes. “It’s not producing as much emissions.”

The Met Office’s collaboration with Microsoft further reinforces this sustainability story. The new supercomputing capability is powered by 100% renewable energy and designed to be among the world’s most energy‑efficient systems. According to the Met Office and Microsoft, the platform is expected to save more than 7,000 tonnes of CO₂ in its first operational year alone.

Blending physics and AI to maintain trust

For all the promise of AI, Kirstine is clear that machine learning will not replace physics‑based models in the foreseeable future. Her vision is a blended system that uses both approaches in combination, exploiting their respective strengths.

 “The future is going to be a blended approach that draws on the strengths of both,” she explains.

Physics‑based models provide an interpretable framework grounded in the laws of nature, helping forecasters understand why the atmosphere behaves as it does. That scientific insight is essential for trust, especially when forecasts inform critical decisions on flood defences, heatwave response and aviation safety. Physics models also generate the synthetic training data required to build robust machine learning systems, especially as climate change makes historical observations less representative of future conditions.

Machine learning models, meanwhile, excel at speed and pattern recognition, capturing subtle relationships in the data that may be difficult to express explicitly in equations. However, they are constrained by their training data and may struggle with so‑called “grey swan” events – rare but physically possible extremes that drive many of the most serious impacts. 

“Those grey swan events, like major heatwaves or extreme flooding, are exactly what we need to predict,” Kirstine says. “That’s why physics still matters.”

AI for Everyone

Embedding AI into forecasting is only part of the transformation. The Met Office has launched a wider change programme called AI for Everyone, designed to bring AI tools and understanding to staff across the organisation. 

“AI for Everyone is a transformational change programme that covers everything that happens in the Met Office,” Kirstine explains.

The programme includes training, clear policies, hackathons and an annual AI for Everyone in the Street event, where teams share how they are applying AI. It spans corporate services, core science and frontline products, ensuring AI is deployed wherever it can add value – from HR and finance to climate downscaling and customer‑facing services.

In corporate functions, the Met Office is rolling out tools such as Microsoft 365 and GitHub Copilot to automate routine tasks and support coding. In scientific research, AI is being used not only for FastNet but also for high‑resolution climate downscaling, supporting both mitigation planning and adaptation decisions. In products and services, teams are exploring large language models to communicate complex forecasts more clearly and deliver tailored information in formats decision‑makers can use.

“We’re asking how we can get the right information, at the right time, in the right format, to the people who need it,” Kirstine says.

Diversity as a catalyst for innovation

For Kirstine, the AI transition is inseparable from the question of who gets to build and govern these systems. She cites the Lovelace report, which found that only around 20% of the UK tech and data science workforce are women, echoing similar findings from the Alan Turing Institute. 

“That’s really not good enough,” Kirstine stresses. “The needle hasn’t moved and that should worry all of us.

“Diversity is a catalyst for innovation. And this is exactly the moment when we need to be at our most innovative.” 

For the Met Office, that means treating AI as a team sport, blending disciplines, backgrounds and perspectives to deliver public services that are accurate, trusted and aligned with net zero goals. 

“The future is not that AI is the team,” Kirstine concludes. “The future is AI’s role on the team.”