Technology
Richard Turner
Richard Turner is Research Lead, AI for Weather Prediction, at the Alan Turing Institute, and Professor of Machine Learning in the Department of Engineering at the University of Cambridge. He was a founding Co-Director of the UKRI Centre for Doctoral Training in the Application of Artificial Intelligence to the study of Environmental Risks.
How connected are recent advances in weather forecasting to progress in large language models?
They are very tightly connected although slightly separated. Many of the AI weather forecasting models use transformers. The reason why they are so fast and so capable of learning is because of that transformer architecture and all the work the community has put into optimizing it. But in weather, it is also good to build in some of the physics, like the fact that Earth is a sphere and that weather can go around the world. For this reason, weather models sometimes also use graph neural networks where it is arguably simpler to build in this physical knowledge. Thus, weather models are closely related to large language models, but a bit different in that, for example, they must represent the planet as a sphere.
Compared to previous methods of weather representation, are we seeing an acceleration similar to what large language models have shown across different applications?
The acceleration achieved through AI-based weather models is even greater. In the weather space, people previously used supercomputers and massive physics models, and that’s extremely costly, not just in terms of maintaining the computer but also in terms of maintaining the team to support the software that runs on the computer. You needed tens or hundreds of people to do the job. The machine-learning models are about a thousand to ten thousand times faster than traditional ones, which is of course a huge speed gain, but the software is much simpler and easier to innovate, so you can do that with a team of five to ten people.
One of the problems with LLMs compared to previous approaches is that they use a huge amount of energy, while “weather AI” consumes less energy than previous systems. Is that right?
“Weather AI” will be a thousand to ten thousand times less expensive, also in terms of energy consumption. Of course, AI models come with the cost of training, which previous systems didn’t have, but it turns out that the training cost is not huge. It’s maybe ten thousand GPU hours, whereas modern LLMs require tens of millions of GPU hours for training.
But don’t weather models require a very large amount of data?
Weather data are far heavier than language data: a data point in language may consist of a few pages of text, while a single weather data point can be several gigabytes in size because it represents the entire Earth via a fine grid with many atmospheric layers, with numerous variables such as pressure, wind, temperature, and humidity. But compared to LLMs, weather models have fewer parameters and therefore require fewer GPUs running in parallel to make predictions.
Compared to traditional weather models, we don’t know how “weather AI” models work. Is that a problem?
Yes. It is a bit of a problem. To build faith in a system, you want to know how it works. It’s perhaps not such a big problem as you might expect, because people didn’t know all the details of previous physical models, as they had 9 million lines of code and had been built over 20 years or so. Thus, no one person knows in detail how the physics models work either. Moreover, meteorologists weren’t as worried about how “weather AI” worked as I expected–they are mainly focused on the quality of the forecasts. And the fact is that with this new approach, we can improve both the efficiency and quality of forecasts.
What lies ahead? What are the main issues?
One issue is extreme weather. This is where accurate prediction is most critical. Here, it is a bit of a mixed picture. Machine models are better at tracking hurricanes and tropical cyclones, for example, but are worse at measuring the intensity, at predicting how strong a storm is. AI models have tended to underestimate the strength of storms. Additionally, there are encouraging results in extreme rainfall and floods. In Dubai last year, there were incredible rainfalls. Nobody had seen anything like that there, and there wasn’t anything like that seen in historical records. But the machine-learning models were still quite accurate compared to traditional models, maybe slightly better. One of the rules of thumb has been that machine learning models are OK as long as they have seen similar weather conditions somewhere in the world before. So even if they hadn’t seen that kind of rain in Dubai, they had seen it in Germany, for example. The real problem is when they haven’t seen the extreme case anywhere before. And with climate change, of course, that can be the case in many situations. So one of the main questions in the field is how to address this limitation. Training on data from climate simulators as a form of data augmentation may help.
Is there more to understand if we want to use real-time data?
The future here is a machine-learning model that can ingest data in real time, continually update its forecast as new data comes in, and incorporate the sorts of data that don’t go into traditional weather models: you can imagine flying drones into targeted locations to get hyperlocal forecasts, or directing satellite data about specific regions, or even using the cameras and sensors controlled by organizations willing to collaborate. At the moment, our systems don’t do that. In the longer term, we will have a system that knows where weather is most uncertain and automatically directs instruments – drones, satellites, or weather balloons – to targets in particular locations where maybe extreme weather is going to occur: at the moment there isn’t such a thing as a feedback loop between the forecast and the observation network. That’s what is going to come in the future.
Imminent Research Report 2026
A journey through the next generation of AI —the moment when machines begin to learn from and interact with the real world in real time.
The ultimate resource for understanding the next phase of AI innovation. Written by a global, multidisciplinary community. Designed to navigate a world where machines mature, learn from experience, and retain memory—while humans remain responsible for how AI acts in the world, and language continues to be the primary way intelligence is created.
Get Your Copy NowDoes it need new science?
It does. The science around active sensing has not been fleshed out yet. We started. But we are not done yet. Another exciting thing that is going to happen is the integration of weather foundation models with large language models. We would like to have a user interface that helps people interact with weather forecasts and prepare for upcoming events. Imagine you are in Africa, and they tell you that your farming business is going to be flooded, but you don’t have the means to get financial advice, or insurance, or to make a plan to limit the damage. Or imagine you are a farmer and you are looking for advice about what to plant, when to plant it, when to harvest it, when to spray with pesticides: an LLM that knows the weather very well could help in your language, and in a way that is easy to understand and implement.
There could be a lot of applications for that. Can we think about mobility and many other areas?
We should think about where there is an intense demand for weather forecasting. Energy production and distribution is an important consumer of weather forecasts to plan for when it can be forecasted that there will be less wind or sun. Public health is also a potential customer, for example, to know if there will be a drought or extreme heat. Cities will need forecasts about air pollution. And weather forecasts will be needed in the management of supply chains, from shipping to rails, and all sorts of deliveries. Applications for insurance companies will be also available as well as for financial markets.
Could we also think about using weather models to forecast climate change and prepare for new conditions in the long run?
AI will definitely play a major role in making future climate predictions, but it is not yet clear precisely what role this will be. Machine-learning approaches are very accurate when making predictions in situations that are regularly encountered in the training data. But for climate, this is not the case and AI has to generalize to significantly different situations. We are working on to address this. Initial experiments have been promising, but it is early days.
