Transcribe your podcast
[00:00:00]

We are back with our weekly segment, AI Decoded. Welcome to the program. We have had a summer break from AI Decoded, but if, like me, you were on the British beaches sheltering from the rain, then maybe you were scanning your mobile weather app to see if the sun might ever reappear, which got us thinking, what about AI and the weather? How do you predict climate when it is changing so fast? How do you process that incredible amount of computerized data that is now being generated? Well, you model it, and that is where AI is making huge advances. There is a forecasting revolution underway, so accurate, says the Guardian, and now in much more accessible come out that very soon governments around the world will be able to save lives and protect livelihoods before extreme events even occur. We'll hear from the team at Oxford University who are filling in the gaps with AI and making it more readily available through cloud computers. Computing. Or how about this from the EU, Destination Earth, a digital copy of our planet on which scientists are running complex simulations to predict natural phenomena. Ai, combined with climate science, powered by supercomputers.

[00:01:16]

A digital twin, if you will, that will help scientists predict the evolution of climate change. With me as ever, our regular commentator and colleague, Stephanie Hair is here also in the studio. We have a very well-known meteorologist Florence Rabier, Dr. Rabier, Director General of the European Center for Medium-Range Weather Forecast, and joining us also on Zoom, Professor Stephen Belcher, who is Chief of Science and Technology at the UK Met Office. Welcome to you all. Florence, we're going to start with you and the Earth's digital twin that you and your team have built in collaboration with the AI industry. So let's get a view view for the viewers. Let's just show the viewers what it entails, and we'll talk off the back.

[00:02:06]

To create a better future, we must push the boundaries of today. Simulations of our Earth system, known as digital twins, will help us understand, predict, and plan for a rapidly changing world. This twin will offer highly detailed interactive data that can support decision-making around extreme weather events. This twin will show us possible futures. It simulates different climate change scenarios over many decades, helping us to be ready for whatever happens.

[00:03:07]

Wow. In simple terms, you are stimulating the natural phenomena and the human activities on Earth, putting it all together through this supercomputer, and what? Putting it onto a digital twin of the Earth?

[00:03:25]

Yes. So digital twin of the Earth, because it's this supermodel of everything that the Earth is doing that we can predict through computing equations. So it's a model where all our knowledge of the physics of the atmosphere and the Earth system is encompassed in that model. So it's a computer program that we put on a supercomputer and we run it. But all our knowledge of the physics accumulated since Newton and all is there about gravity, condensation, storms, et cetera. It's a digital twin of the Earth because It's very accurate and it has a very high resolution. Also it's interactive. You can play a bit with it and simulate what-if scenarios.

[00:04:10]

I imagine that in times gone by, you would do that at a very local level. But of course, we're all interconnected. The world is a global environment. Our weather systems and activities are all connected. How does this enable you to improve the forecasting that What we do?

[00:04:31]

At FMWF, what we do, we run this global model. It's really across the whole world. Because if you want to know the weather in Europe now, you have to know what happened in the US a few days ago and in the Atlantic, and even in the Pacific, if you want to predict the weather seven days beforehand. It's all interconnected. You're right, you have to start with a global scale, and then you can refine at the local scale as well. But you really have to know whatever happens on the world at any point in time in order to go further in time in your predictions. That's what we've been doing for about 50 years in collaborations with our member states, 35 countries in Europe supporting this work.

[00:05:12]

Florence, is this a uniquely European initiative or do you work with other partners around the world?

[00:05:17]

There are several global models in the world. Because we are European, we are working in collaboration with 35 meteorological services in broadly Europe. Brilliant. But there's a The rate and model, et cetera.

[00:05:31]

Okay. Are you using historical data or live data or both?

[00:05:37]

Probably both. I mean, for the weather forecast, usually we use the current data, so the data we've seen in the last 12 hours, but the model had seen the data beforehand, and it's a continuous process. We combine physics and data, we go forward, combine again. So the latest forecast uses the latest data. But historically, we've used the data from the decades and rolling like this.

[00:06:02]

I tell you what, we've got a real-life example that mapped the recent typhoon, typhoon gamey, I think it was called in July this year. So you see all the lines around the two main lines. There's the red line and the black line, which we'll talk about in a second. What are the other lines that we're looking at?

[00:06:20]

That's typically what we do as a forecast. We predict the weather, but in particular, we concentrate on severe events like that because this typhoon gave me really had dire consequences with 100 people dead and millions affected. So what we produce every day, we produce not just one forecast, but we produce actually several forecasts together to depict the whole probabilities of what the weather will do. This way, we don't just simulate the track of the typhoon, but all the possible tracks that we think the typhoon will take in the next few days, these are the gray lines.

[00:06:58]

So which is the AI model?

[00:07:00]

The gray lines are our physics-based model, and the black line is the real observed track of the typhoon. In the blue, what you have is our best estimate that we had before AI of where the typhoon would go and would hit China.

[00:07:16]

Are you telling me that red one is the AI? That's almost tracking exactly what happened in real-time.

[00:07:21]

In that case, it is. That's incredible. It is incredible, but you can't judge everything on one case, of course. We accumulate all these cases and we do statistics, but it is true that the AI models are in general about 25% more accurate in predicting the track of tropical cyclones, typhoon, and hurricanes, which is huge, but they are not doing everything right either. So in particular, in terms of the intensity of the typhoon, they are actually about 20% worse. So it's not all perfect.

[00:07:54]

You've got a real interest in this because I know that you grew up in Tornado Valley in America, right?

[00:07:59]

Yeah. Tornado Alley is the midwestern part of the United States that runs from North Dakota all the way down to Texas, and then probably for 500 to 1,000 miles on either side. I grew up just outside Chicago, and routinely, we practice these drills as children. You get a little bit of warning. We're talking seconds, and you have to find the nearest basement and get underground because a storm will come through that can destroy an entire town in seconds.

[00:08:24]

It's quite terrifying. Presumably, this could tell you which street to go to.

[00:08:28]

Well, we're not at that scale, especially as we work.

[00:08:31]

I have big expectations for it. But it's getting better and better all the time. It is getting better all the time. Let's bring in Professor Stephen Belcher, who is the Chief of Science and Technology at the UK Met Office. Welcome to you. So far, we've talked about global weather patterns, Stephen. Climate trends, mapping, evolution of weather patterns. But how much more precise is weather prediction getting day-to-day because of this AI technology?

[00:08:58]

Well, it's worth remembering that There's always pressure to increase the accuracy and the utility of weather forecasts. Today's a great example. We've had certainly lots of rain here down in the southwest of England. With climate change making extreme events even more extreme, we're demanding that our forecasts get better to help us understand what the impacts of those might be. Also, we've got new applications of weather forecasts. Just think about the rollout of renewables. This is meaning the weather is now the the fuel of the future. So understanding that fuel is another application of our weather forecasts. And so to make them more accurate, we need increased lead time, so we need warnings further ahead of when we're getting these extreme events. But we'd also like finer detail, just as Florence was describing earlier.

[00:09:50]

I was making quite big demands of Florence, but I'm going to put up an image here about just how accurate this can get. So here's an image of London that we'll all recognize. You can see the Millennium Dome there, the bend in the river, East London. Tell us what we're looking at here on the left and on the right.

[00:10:09]

At the Met Office, we complement what's done in Florence's organization, the European Center, by producing high-resolution, physics-based modeling of the weather over the UK. The left-hand side is showing you the grid that we divide London into in order to provide that weather. What we're seeing now is the weather at that resolution. About one and a half kilometers, we increment the differences in the rainfall and the temperatures. What we've been able to do, in fact, one of our rising stars, Louis Blun, here at the Met Office, working with students at the University of Reading and also at the Bureau of Meteorology in Australia, has devised a machine learning technique to add fine detail onto those routine forecasts that gives us detail at resolutions of hundreds of meters in the temperature.

[00:11:06]

You can tell the temperature and the heat and the rainfall literally over the Millennium Dome.

[00:11:11]

The temperature at this stage and maybe other variables in the future, but the temperature is what Lewis and his colleagues worked on. The reason this is important is that we've known for some time that when we have heat waves, those heat waves are more extreme in urban areas. In particular, those who live in cities will have noticed that the temperatures don't cool down so much at night, and we call that the urban heat island. What Lewis and his team did was to take data, actually from crowd-sourced data in back gardens and citizens in London, their data. A variable quality, frankly, plus five professional weather stations, they mash that together with machine learning tools and augment the regular forecast that we produce here at the Met Office and can then produce these temperature forecasts at these very fine levels. So it's another example of how AI can really change what we're doing in the weather world.

[00:12:16]

Professor Belcher, I've got a question for you exactly about this, taking data from your back garden. Sometimes when I'm standing in London, I will consult the Met Office app, regularly, in fact, and it will say that it's sunny and I'm being rained on. Why is that happening? And second, how or when will I be able to send data to you saying, No, no, here in Hackney, it's raining.

[00:12:41]

You can send data to us right now. It's called weather on the web, the WOW site. So please do that. You can launch your longitude and latitude and send in your data to us. And as I say, that's the data that Lewis and his colleagues used to produce this. I think in In terms of weather forecast, let's not forget that over the last 50 years, through the advent of satellite observations and other observations right around the world, the improvement of those global models that Florence was describing earlier and the increase At the scale of supercomputers that we've got, these physics-based models we've got, the weather forecast has improved tremendously well. One of the statistics we describe is that the weather forecast improves by one day per decade. So the four-day forecast now is as good as the three-day forecast was 10 years ago. So this is often referred to as the quiet revolution in weather forecasting. What AI is doing is really accelerating that revolution. So it's a loud revolution.

[00:13:47]

We're going to continue the conversation. We have to get a short break, but we're going to see how this can be applied around the world. Coming up after the break, we'll bring in the climate physicist Dr. Shruti Nath. She's part of the physics team at Oxford, where they've just pioneered a new approach to predicting extreme with us. Stay with us. Welcome back. We are warned repeatedly that climate change will affect millions of people worldwide. In fact, it's already affecting lives and livelihoods, and particularly so in some of the poorest regions of the world, where they don't have access to this real-time forecasting or the vast computer power needed to produce it. Dr. Shruti Nath is a climate scientist at Oxford University. She's been working with the UN World Food Program to develop an AI system that is pulling together all this data, current and historic, and applying that to localized area. That information can now be condensed and shared through cloud computing to help the governments and aid agencies better prepare for climate disasters. Let's bring in our guest then, Dr. Shruti Dhaf. It's good to talk to you. Thank you very much for coming on the program.

[00:14:53]

There are some brilliant physicists like you in the Oxford University Department. I want to better understand, though, what the AI is doing to speed up the process and fill in the gaps.

[00:15:05]

Yeah. So thank you for having me. At Oxford Physics, we're exploring hybrid modeling approaches. So we're looking at how AI can best complement our existing physical weather models. So as Florence says, these models have all the physical knowledge that we've accumulated since Newton, and we're complementing it with AI. Particularly at Oxford, where We're looking at rainfall since this is a high impact, very localized feature of the weather. What we're seeing is that when we take the best of the physical weather forecasts, we can really use AI as a data-driven technique to correct the structural errors that exist in these physical forecasts that could arise from incomplete representation of the atmospheric processes to better deliver actual, accurate rainfall forecasts within the region that we work in. So that's the greater Horn of Africa.

[00:16:05]

Dr. Naaz, how do you see ordinary people in the regions where you're working being able to access this very sophisticated and high power technology that you're working with?

[00:16:17]

That's a very good question, and that's actually, in my opinion, one of the real strengths of AI. It's a very low cost, lightweight model of being able to represent very complex phenomena. We work closely with all the local meteorological bodies, and we work with developing the model with them. They actually run the AI models in-house, and that means that they can actually generate weather forecasts on a laptop. Mind you, that's a laptop as compared to a supercomputer, which is what typically is used to generate weather forecasts.

[00:16:54]

Can you give us an example of where you've used that? A real-time example.

[00:16:58]

Of course, yeah. We use it in Kenya. They update the forecast every day in Kenya and Ethiopia. The forecasts are also available on a website. The website name is cgan. Ikpak. Net, and they're updated every day from the in-house forecast generated on their equipment. It really is a way of giving these people a bit more accessible weather forecast Forecasting.

[00:17:30]

But presumably, the breakthrough of all this is that if you know what's coming and the long-range forecasting improves, the aid agencies can store forward the aid that they're going to need for what's coming at them. So often on our programs, we're saying, Well, we can't get to these inaccessible areas, but now the stuff will already be there because we've already forecasted what's coming. Exactly.

[00:18:00]

We particularly actually focus on linking research to action. We work with linking these forecasts to anticipatory action. As you said, we can have these long-range forecasts and also chipping in on what Florence mentioned about how weather is quite chaotic. There's lots of possibilities that can arise from a given starting point. We have a lot of uncertainty, and you need to actually generate a lot of different weather forecasts to explore that. Ai AI allows you to do that in a very low cost way. You can generate forecasts that explore the uncertainty space in a very low cost manner so that you can actually properly inform anticipate reaction in these areas.

[00:18:45]

We've talked a lot about weather. What we've not actually talked a lot about is climate change. Of course, there are climate deniers out there, Florence, that we must acknowledge. I'm going to put a picture on screen. Do you ever remember This was a tornado that was coming at the Florida Panhandle. Also, there were some questions about whether it might go to Alabama. They got a Sharpie out and they actually drew it on the end, the Trump administration, which tells you that it clearly is something that people try to play with when we talk about the way climate is changing and what weather is going to do. But your forecast is so accurate, Florence, and presumably with your digital twin Earth, you can predict how climate is going to evolve well into the future.

[00:19:38]

Well, exactly. We use the same weather model to do climate models as well. They are just a bit more complicated, but it's based on the same modeling. But also we can document climate change, and that's what we're doing with the Copernicus program from the EU, going back from 1940 and really depicting predicting what the weather and climate have been doing every hour. From 1940 to now. We have this picture of the Earth, and we can then document how much the temperature have increased, how much the frequency of storms have increased, et cetera. So it's predicting it, but there is already this reality. We have enough information to know what has had happened already. And then with these models, then we can do a digital twin of the climate as well and go forward in the in the future, with different scenarios, of course, of what will happen in the reduction of greenhouse gasses, because, of course, it all depends how much we can reduce the amount of carbon dioxide in particular that we put in the atmosphere.

[00:20:45]

I've got a question for all three of our distinguished scientists.

[00:20:48]

We've only got a minute left, so you're going to have to make it quick.

[00:20:50]

Go on. We keep hearing that we're running out of data and that this is a big problem for AI. But I wonder if that's actually true, particularly when it comes to weather, climate change, and biodiversity loss. We have to fix, obviously, the climate change and biodiversity problem. Do you feel that there's a way for citizen scientists to get back into action and be submitting data to all of you scientists so that you can help us fight these bigger problems?

[00:21:15]

Stephen, pick that up because we've just about 30 seconds left.

[00:21:19]

Yeah, it's a great shout. As we talked about earlier, we've got the weather on the web. There's another great crowdsourcing initiative to look at budding and early sighting of insects around the UK, which we've also connected with climate change here at the Met Office, along with many other partner organizations. I think it's a great shout for Citizens Science, this one.

[00:21:42]

I could talk plenty more, as I always could on this program every week. We never get to the bottom of everything. But listen, Florence, Steven, Dr. Nath, Stephanie, thank you all very much for your time. Really fascinating discussion. Just a reminder, we are putting all these half-hour programs on the BBC's AI decoded YouTube site, so you can find all our past programs there. We'll do it again the same time next week. Thanks for watching.