Quiet AI revolution taking weather forecasting by storm

In recent weeks, we’ve had some exciting announcements about the benefits of AI in weather forecasting. Some of the world’s leading tech firms - such as Microsoft, Google and China’s Huawei - are all getting in on the act. So, what do we need to know about it and how it works? And - crucially - will it lead to more accurate forecasts? Richard Turner is a professor of machine learning at the University of Cambridge. He’s also a visiting researcher at Microsoft and busy working on developing what he says are far more energy efficient and more accurate AI systems for telling us when it’s finally not going to rain all over the UK…

Richard - The traditional approach to weather forecasting involves two steps. First, observational measurements from satellites and weather stations, and you combine them all together and get an estimate for what the current state of the atmosphere is. In the second step, having figured out what the atmosphere looks like at the moment, you run a big simulation on a supercomputer that evolves the atmosphere and tells you what it's going to be doing, say, a day ahead or a week ahead, or two weeks ahead. To date, AI has revolutionised that second step, the forecasting step, and it replaces that with a machine learning system. The great advantage of the machine learning system is firstly it's about 10,000 times cheaper, computationally. You can literally run these things on a big desktop computer. Secondly, you can, in certain circumstances, actually get a more accurate, as well as a much cheaper system.

Chris - And how much better is it?

Richard - That's a really hard question to answer, but the improvements are of the order of maybe 5%. Just to put that into context, when you take a national weather forecasting agency's predictions, it might take them a year and a large team and a big supercomputer to get an equivalent improvement. Those improvements are really tangible to farmers in agriculture, or if you are using these things to predict renewables for energy production, it's a big deal with these sorts of numbers.

Chris - Where's the bottleneck now then?

Richard - One of the current bottlenecks is actually you still need to run a conventional weather forecasting system in order to get that initial estimate for the state of the atmosphere. One of the things I'm really excited about in my group, we have a project called Project Aardvark, where we're looking to replace the entire system with a machine learned variant. The system ingests data from weather balloons and satellites, weather station measurements, and then directly outputs a forecast for the next 10 days. It's about as good as the best systems from the US and a little bit behind the best systems in Europe.

Chris - And how do you train something like that? Was it literally you took enormous amounts of climate data, or initially not climate data, just data, and then honed in on climate data and then said, this is the outcome, this was the rainfall, this was the wind direction, this was the temperature. Is that how you trained it?

Richard - Yeah, so in the Aardvark system, that is pretty much how we trained it. We trained it over about 10 years of data.

Chris - One of the constraints of the present system is that it's down to how good the data are in your particular geography. So in the UK people are saying we are predicting weather with some accuracy down to sort of the square kilometre, aren't they? Whereas, you go to the middle of nowhere, there's much less data, much less accuracy. Now you can't improve on that presumably?

Richard - No, that's right. You're fundamentally limited by the accuracy and so I think there are real challenges here. For instance, underdeveloped nations, the weather forecasting is much poorer and more limited, partly because of the lack of observational measurements. But one of the great advantages of the Aardvark Project is it's easy for developing nations to develop personalised weather forecasting systems, which previously would require huge teams to operate, deploy, and maintain. So we're trying to figure out a way to get Aardvark to those nations and already seeing it can lead to quite big benefits over the traditional approaches.

Chris - Would that be almost like an open source system, and they just have a big desktop and they can just run this at home, program it with local data and they've got their own bespoke local weather forecast?

Richard - That's right. They would need a bit of computational resource to train it, but once it's been trained - and it isn't terribly expensive to train compared to having a supercomputer - they literally could run it on a desktop or a laptop computer and take ownership of that model.

Chris - It sounds pretty revolutionary. Why have I not heard people shouting from the rooftops about this for much longer until I met you?

Richard - Well, I think if you've looked for the signs of it, you'd have seen it. But it's been called the quiet AI revolution by the chief scientist at the Met office. It's absolutely ripping up and transforming the field and the big meteorological agencies are now hiring big machine learning teams. It's not clear what the future is going to be and exactly where AI will be embedded into the current processes, but it's clear it's going to make a massive difference.

Chris - One of the things that's emerged from a lot of these large language models that's got people spooked is - it's more accurately termed confabulation but the wider public refer to hallucination - it just makes stuff up. So is it going to do a sort of reverse Michael Fish and say, 'there's going to be a hurricane!' In fact, there isn't.

Richard - I think that is a great question. One of the things that we need to do is rigorous evaluation of these models and one of the great fears is when you need it most in severe weather, these things will confabulate. That of course would be really, really bad. We know machine learning systems do worse in the extremes where they don't have much training data. I think that's the particular challenge in the climate case as well because obviously we don't have much data about how the climate's going to change. So there has to be a few years I think of assessment of these models and looking at how well they hold up in extreme events. The early indications are that they hold up much better than people expected and I think even the meteorologists have been surprised that they still seem on par with the best numerical systems in these extreme events.