A sea-change is coming for weather forecasting. Well-known as a complex and labour-intensive process reliant on satellites, sensors and supercomputers, the industry is being transformed as AI and machine learning enable more accurate and timely forecasts than ever before.
The Met Office is at the forefront of this shift, blending AI with traditional physics-based models to deliver richer insights, greater computational efficiency and timelier predictions. Kirstine Dale, Chief AI Officer at the organisation, describes the shift as a “raucous revolution”, with machine learning models now running tens of thousands of times faster than conventional methods.
Yet, according to Dale, AI is not a silver bullet – and human input is still needed to interpret unprecedented events, validate results and maintain the public trust built over decades. In this interview, edited for length and clarity, Dale explains how the Met Office is integrating AI into its operations, the opportunities it offers and the need for human input amidst the technological revolution.
How does the Met Office grapple with AI and machine learning?
The Met Office is embedding AI across its end-to-end functions, helping realise the value of this breakthrough technology in terms of weather and climate science and services.
Over the past few years, we’ve specifically been integrating AI across the forecasting chain. For example, we collaborated with DeepMind to produce a deep-learning system that turns radar data into highly accurate short-term rainfall forecasts. We’re also using machine learning to develop ‘emulators’ to replace some of the most computationally expensive physics in our weather models, as well as for bias correction and to improve site-specific temperature forecasts.
While it wasn’t called AI at the time, we’ve also been using Gaussian processes for downscaling, clustering, cloud classification, etcetera, for a really long time. All that’s to say that data science is not new for the Met Office. We aren’t coming to AI from scratch, and that’s enabled us to accelerate its deployment pretty fast.
What makes AI such a game-changer for weather forecasting?
The Met Office is essentially a big data organisation. Weather observations are collected around the clock worldwide, combining measurements of key variables with satellite imagery to build a picture of what’s happening in the atmosphere.
We have roughly 215bn observations coming through the organisation every day, we run 3m lines of code and produce 18 terabytes of data each day. This is a phenomenal amount of data flowing through the organisation. So we were really delighted and excited when the AI revolution started, and when we could see some potential for how we might tap into the tools and techniques that were coming out of the AI revolution for extracting value and meaning from data sets.
Is it predominantly in terms of accuracy that you see AI making a difference, or are there other benefits?
Accuracy is definitely a part of it – machine learning models only really started becoming competitive at the end of 2022, and now they’re matching (and sometimes surpassing) our physics-based approaches when it comes to accuracy. But it’s not the only benefit. Another real game-changer is speed.
These models aren’t just a bit faster, they can be tens of thousands of times faster than traditional physics-based models. That matters for two reasons. First, because they run so quickly, you can generate forecasts much closer to the moment you need them. Second, they’re computationally light. Training a model is still expensive, but once trained, the inference phase (running the forecast) requires relatively little compute. That in turn reduces the environmental footprint of modelling, and opens up the possibility of running forecasts on smaller systems. In the future, you could imagine forecasts being generated on laptops or even smartphones, which would massively increase accessibility.
To put this in context, forecast accuracy has historically improved by about one day per decade, something that’s often called the “quiet revolution.” AI has the potential to accelerate that curve dramatically. It could deliver a real step change, what you might call a “raucous revolution” compared to the steady gains of the past.
Where do you see the limits of AI in weather forecasting?
AI will not replace physics-based forecasting. The physics models are vital for explaining why a forecast is unfolding in a certain way, which is essential for trust, and AI models are limited by the data they’ve been trained on. They may struggle with unprecedented events – volcanic eruptions, for example, or extremes that haven’t yet appeared in the training data. And, in a changing climate, that’s a real concern. The climate of the future won’t look like the climate of the past, so there are limitations to machine learning models trained using historic weather.
In practice, we see AI and physics-based models running side by side: physics for insight and trust, AI for speed and efficiency. Used together, they could make forecasting both more powerful and more accessible than ever before.
What other challenges have you seen in terms of rolling out AI?
There are a lot, to be honest, because this is still something relatively new and people are still determining how best to use it. The main one, though, is trust. We need to have complete confidence in the machine learning models’ ability to perform well and accurately, and be competitive with the physics-based models, before we would even consider rolling it out.
For us, a big challenge is ensuring we trust a model before we share it, or else we undermine the public trust we’ve spent the past 150 years building.
How can you ensure this trust is there?
Over time, we’ve developed rigorous procedures for evaluating the performance of new systems. That framework now applies to AI as well. We won’t operationalise a machine learning model until it has been thoroughly tested against strict performance criteria. The process isn’t starting from scratch; we’re building on decades of experience in testing, validating, and improving models, and that gives us confidence that when we do integrate AI, we can trust the results.
Looking ahead, do you think we’re just at the beginning of a paradigm shift around AI in weather forecasting?
Yes, I think we are. Machine learning approaches for weather prediction are still being developed, and while they hold real potential, the shift hasn’t fully happened yet. I often think about this through the lens of Amara’s Law: we tend to overestimate the impact of new technologies in the short term and underestimate their impact in the long term. That feels true of AI.
Early on, some people imagined avatars at the Met Office doing the forecasting, which sounded fun but was clearly an overestimation of what was immediately possible. What’s more likely is that, as a global community, we’re underestimating how transformative AI will be over the longer term.
That said, our expert human forecasters will remain central. Their roles may evolve, and they’ll have more tools at their disposal, but their judgment and experience will be just as critical as ever.
Read more: AI may be changing cybercrime but businesses can still fight back, says Experian’s Christine Foster
