// This is a machine generated transcript. Please report any transcription errors to will-help@illinois.edu.

[00:00:00]
spk_0: The weather and climate often cause concern and curiosity, but understanding the facts behind the phenomena can be a challenge. We're here to help. From the University of Illinois, Urbana-Champaign and Illinois Public Media, welcome to Weather Realness, a weekly podcast about the skies above Illinois and the world.

[00:00:30]
Steve Nesbitt: I'm your host Steve Nesbitt, head of the Department of Climate, Meteorology and Atmospheric Science at the University of Illinois Urbana-Champaign.

[00:00:37]
Jeff Frame: and I'm Jeff Frame, also a professor in the Department of Climate, Meteorology and Atmospheric Sciences at the [U of I].

[00:00:44]
Steve Nesbitt: AI now has a sizable presence in many areas of our lives from ChatGPT to huge data centers to phone calls and emails that just don't sound like a person.

[00:00:53]
Jeff Frame: However, meteorologists are using AI to improve weather forecasting. For

[00:00:58]
Speaker 3: decades, meteorologists relied on the European Centre for Medium-Range Weather Forecasts, using data from satellites, buoys, and weather stations. Last year, tech company Huawei Pangu boasted their AI weather model that could produce a weeklong forecast in 1.4 seconds, compared to the old physics-based model that could produce one forecast every six hours.

[00:01:20]
Steve Nesbitt: To learn more about how AI is used in weather prediction, we're joined by Dr. Randy Chase, atmospheric data scientist for Tomorrow.io, an AI weather forecast group and graduate of the Department of Climate, Meteorology and Atmospheric Sciences at the University of Illinois. Thank you for joining us, Randy.

[00:01:37]
Randy Chase: Thanks for having me.

[00:01:39]
Steve Nesbitt: So Randy, can you tell us a little bit about your work with Tomorrow.io?

[00:01:42]
Randy Chase: So Tomorrow.io is a startup weather company providing forecasts for the private industry, and we do that through three main pillars of weather forecasting. So the first part about that is we collect novel microwave sounder data. So this is satellite data that is being collected in the microwave spectrum that helps provide information on temperature and humidity. And we first started launching those in about 2024 and have since launched 11 total sounders. So this is our first part of the business and I'm involved with preparing that data and doing science with the microwave sounder data. 

Beyond that, the second part of the company is building more specialized forecasts for customer needs. So there are a lot of public weather forecasts out there such as the ones from NOAA that a lot of people use, and they're great for general use, but there's a lot of commercial need for better results for things like precipitation down with timing down to like 15 minutes or hitting a specific location, as well as things like when will my weather station or my factory observe freezing temperatures type of perspective and for how long. So, at Tomorrow.io we build custom forecasts to help narrow in those broad general forecasts from NOAA and help our customers reduce losses and create more optimal conditions for their businesses. 

And then the final part of our company is that we use a GUI or a front-end platform that allows for customers to choose their own thresholds and build their own alerting system. So essentially they can have automated alerts that go across their entire company.

[00:03:18]
Steve Nesbitt: Randy, let's talk a little bit broader about AI because it's basically infiltrated almost every part of our life and you can't talk to anybody anymore without talking about AI. But of course there's AI like ChatGPT, there's AI that we see on social media that everybody laughs at. But can you talk a little bit about, you know, broader about AI and then later on we'll talk about how it affects weather forecasting.

[00:03:45]
Randy Chase: Yeah, so artificial intelligence or AI is basically a computing system that does tasks that would normally require a human, and maybe that definition needs to change as we become more and more incorporated with artificial intelligence in our everyday life. So that's kind of the broad set of artificial intelligence and something that we use more regularly in atmospheric sciences and weather prediction is more called machine learning. And this is an algorithm that learns from past data and does automated decisions based on that learning from past data.

[00:04:19]
Steve Nesbitt: So Jeff, you've been doing weather forecasting for a long time. What is the experience of and the procedure for doing weather forecasts without AI?

[00:04:28]
Jeff Frame: Yeah, so very short term forecasts. We can look at observations, for example, look at radar to issue a tornado warning, something like that. But if we're going out, you know, several hours or certainly beyond a couple of days, those forecasts are almost entirely based on computer forecast models. The first one of these was developed all the way back in the 1950s and then they've been refined since over the decades. 

But basically the way these models work is imagine a sheet of graph paper stretched around the world and everywhere the grid lines on the graph paper intersect, these computer models solve a very complicated set of atmospheric equations, and these equations are so complex that you know even Isaac Newton or Albert Einstein can't pick up a pen and write down, oh, X equals this and solve the equation. But of course the atmosphere is three dimensional. So we have not only one sheet of graph paper but maybe 50 or 60 sheets of graph paper stacked up stretched around the world, and it takes a lot of computing power to run these traditional, we call them numerical weather prediction models or NWP for short, such that these models take anywhere from a couple to several hours to run to produce like a two week forecast on a supercomputer. 

so if you went and tried to run, you know, the American long range forecast model, for example, it's called the GFS on your computer, I mean it would take, it would take many, many days to do that and obviously you wouldn't have the computing power to get a forecast out of that. You would get a hindcast.

[00:06:10]
Steve Nesbitt: So when we're talking about running an AI weather forecast model, we're not talking about anything like artificial intelligence from the fact that we're inventing things or making novel things out of nowhere. This is just about learning from what we've seen in the past.

[00:06:29]
Randy Chase: Correct, yeah, and just like how Gemini and ChatGPT have learned how to generate images of cats and videos of cats doing fun things, it can learn how to evolve a weather system. So that's what these AI weather prediction models largely do is they've watched millions and millions of weather scenes, and they've watched this from something called the [ERA5] reanalysis, which is our long-term best guess estimate of what the atmosphere has been in the past, and it's 50 years or so of weather movies if you want to think about it. 

And we take these weather movies and we show how these weather movies evolve over six hours, and the AI does this millions and millions of times and it slowly learns this approximation for the complex equations. So what this enables is that instead of needing a supercomputer, you can now run a full global weather model. Granted, it's a little bit of a coarse weather model at the moment, but I'm sure that the higher resolution models are coming. We can run a global weather model even on a common MacBook Air. I've done it actually on my own MacBook Air, and it's taken about 10 minutes to run, which is really quite an impressive feat if you've known how much computing power that these weather forecasts used to take.

[00:07:47]
Jeff Frame: So other than producing forecasts a lot more quickly, how else might AI improve weather forecasting?

[00:07:55]
Randy Chase: So I think one of the things that makes me really excited about AI for weather forecasting is AI's natural ability to handle large amounts of data. So something that Jeff didn't quite mention in the weather forecasting [pipeline] is that we need to create a current weather state to get that weather forecast. And knowing the best current weather state will then enable a better weather forecast. So if you know where the weather is currently, you'll know, you'll start to know where it is moving or developing as it goes forward into the future. 

So this is called data [assimilation] and we take large amounts of data, we take satellite data, radar data, and we put them like he was saying on this common grid. But the problem is, is we have to make a choice to prevent this from taking too long to compute and like Jeff was saying, you don't want a hindcast, you want a forecast. So you have to choose how much data you can use in that forecast. But the thing with AI is because it could speed up, you can then potentially run this in a much quicker cadence as well as pace, and then you could potentially handle more and more amounts of data. So that's the part of the science that hasn't quite happened yet, but I'm excited for the field to go that direction to be able to leverage much more of our current observing system, like the satellite data into the weather forecast and hopefully making them better and better forecasts.

[00:09:19]
Steve Nesbitt: And of course, the critical weather forecast that we receive, for example, here on WILL from Andrew Pritchard, Jeff Frame on social media, those that work at the National Weather Service for critical life saving information, tornado warnings and other things. Are those impacted by AI?

[00:09:37]
Randy Chase: I would say that they're impacted by AI in the sense that the weather forecasts who are implementing or disseminating those warnings and watches are using more and more AI tools to help them be better at implementing those forecasts. There's usually an annual weather test bed that occurs down in Oklahoma every year where they're testing out these new AI weather tools every year. And they're using the best ones that come out of that test bed to help improve the weather forecasters. So

[00:10:07]
Jeff Frame: Randy, what do you think about the future of AI and weather prediction?

[00:10:11]
Randy Chase: I think it's an exciting realm to get in, to be in, and so it really has lowered the barrier to do a lot of cool things in the weather forecasting space. In the past, I mean, even just five years ago, a private company like my own at Tomorrow, you would be running these physics-based weather models and as we've talked about already, these physics-based weather models cost a lot of compute power. So the ability for a private company to provide a solution to whoever wants, whoever needs a weather solution, which is a lot of people out there, was quite limited because we could only have so much money to spend on creating these weather forecasts.

[00:10:56]
Steve Nesbitt: Randy, thank you for joining us. Thanks again. Randy Chase is an atmospheric data scientist for Tomorrow.io, an AI weather forecasting group. If you're just tuning in, you're listening to Weather Realness, the weekly podcast about weather and climate. Today we are talking about AI and weather and climate prediction with Randy Chase, atmospheric data scientist with Tomorrow.io, and Gan Zhang, assistant professor in the Department of Climate, Meteorology and Atmospheric Sciences at the University of Illinois Urbana-Champaign. Now let's switch gears to talk about AI and longer range forecasts.

[00:11:31]
Jeff Frame: Meteorologists not only forecast weather conditions over the coming days, but they also predict what weather conditions might be like over the coming months. For example, we can anticipate if large scale trends like La Niña will continue.

[00:11:46]
Steve Nesbitt: Here to help us break down AI's impact on climate prediction is Gan Zhang, assistant professor in the Department of Climate, Meteorology and Atmospheric Sciences at the University of Illinois Urbana-Champaign. Thank you for joining us, [Gan].

[00:11:58]
Gan Zhang: Good to be here.

[00:11:59]
Jeff Frame: All right, before we get started when we're talking about climate prediction in this case, we're talking about climate predictions over the scales of months to maybe a year or so. And these are different from the much longer range climate projections that are used, for example, to anticipate the effects of climate change, how much global temperatures might rise by the end of the century, sea level rise, and those things. So when we're talking about climate predictions today, we're talking about that on the range of months, not on the range of decades. So now that we have that out of the way, Gan, can you tell us why seasonal and long range forecasts are important?

[00:12:39]
Gan Zhang: Yeah, sure. we do these predictions for a number of variables, temperature, wind, rainfall. So we try to figure out how they might look like in a few months to maybe several decades. So one straightforward example is from poetry, if winter comes, can spring be far behind. So that's a prediction based on natural [drivers] like the earth's orbit variations. we also make predictions based on the human [drivers] as well. 

So we make [traditions], we use a [tradition] to do planning such as planting seeds, plan trips or other activities. In recent years, people also increasingly use the [traditions] for financial transactions. For example, we can make informed decisions [on optimized] plans and actions if [climate] models predict a longer [drought] trend in the southwest U.S. then the water resource management can do planning and other remedies accordingly ahead. And also if we have an upcoming El Niño [that] makes hazards more active in our region, then the pricing of the coffee beans and the insurance could change. So that will [affect] our daily life, especially the affordability issues.

[00:14:02]
Steve Nesbitt: [Gan], you're one of our esteemed graduates of our department and you did climate predictions as a student. How did you do this before AI?

[00:14:11]
Gan Zhang: Before the age of AI we mainly rely on two types of models. One is the statistical model, relatively simple regressions, trying to establish the connection between features like El Niño and the Atlantic hurricane activity. That's one of the methods. The other method to work with dynamic climate models. So we put in this initial values [of observing] data into the model, then the model try to work with this grids, it's a bit like the Minecraft game, you have this little boxes represent the earth system, the land, ocean, atmosphere, you solve mathematical equations. The solution needs to be derived from the large computers because it's very complex process. In the end, the model will output some variables that you can work with. So that includes variables like temperature, precipitation and also wind. So this is how the dynamical [predictions] are working in the past.

[00:15:18]
Jeff Frame: Gan, how is AI used in climate prediction?

[00:15:21]
Gan Zhang: For AI's application climate prediction, there are many different types. So right now there are several emerging techniques coming up. So we have some very interesting synergy between AI and climate prediction. So AI tools can learn from vast amount of historical data on satellite imagery and the output from traditional physical models. So I'll give you several examples here. 

the first one is trying to go for the speed. So if you train the AI models using [reanalysis] data or [climate model] data, they can recognize the patterns in the models. And once they are [trained], AI emulator or the AI model can generate forecast very very fast. That means finish months or decades simulation very quickly at a relatively low cost. So that's beneficial. 

Then another venue is to do this process called downscaling. So that's trying to take the low resolution output from physical models and try to increase its resolution to high resolution, maybe city level details. So this kind of application is helpful if you care about what might happen in Champaign city or what could occur to Chicago. So it's crucial to link [climate] information climate information to the neighborhood and stakeholders.

[00:16:53]
Steve Nesbitt: As we're getting into these longer range time forecasts, what are some open questions in the science of climate prediction and how can AI help fill those gaps?

[00:17:04]
Gan Zhang: Yeah, sure. So these are great questions. for the long range prediction, I have some personal niche, you probably can guess I like to think about how to make prediction for extreme events. This can be hurricanes and droughts or wildfires. So for these emerging applications, they are apparently important because they have big impacts on society. So people have been working on these questions for different time periods for hurricanes, people have been working on it for decades, and for wildfire it was relatively recent development. 

So there's a big gap between our conventional tools and the [pressing] societal needs. For example, the [output of climate] models, they come at the very [coarse] resolution. They do not naturally simulate hurricanes or wildfire that well. We also care about the probability of these extreme events. for example, we want to know what's the [chance] of seeing wildfire in the next decades, whether it's going to be 1% or 10%. So the typical way to run many simulations to figure out how many members in these parallel simulations have the signal. if you work with physical models, the simulation can be fairly slow. It's very hard to tease out the right probability. 

So the AI models because [of] their speed, we can potentially run many parallel simulations generate this large ensemble simulation, trying to figure out the tail events with potentially higher accuracy. Of course, when the models agree, that's great. but sometimes the physical model and AI models, [they] can have a very different ideas about the extreme events, figuring out why the models agree [or] disagree will be valuable. So that's a very interesting scientific problem as well.

[00:19:09]
Jeff Frame: AI has a large carbon footprint and can also be water intensive in terms of cooling those data centers. What are your thoughts on this since it's being used for climate prediction?

[00:19:20]
Gan Zhang: Well, that's a great question. So, they do have some carbon water footprint, but if you compare the energy, the water used by running AI models for weather and climate, these are very small compared to these larger language models. Basically the model that power ChatGPT and other applications. So that's one thing, but I do see some interesting synergy between the increasing energy demand by data centers. 

So the carbon and water footprint for these data centers, they are huge, but they can be partly mitigated by the renewable energy, for example, the data center owners, they might own, invest or do business with wind or solar farms. So the energy and data infrastructure depend on weather and climate variations. You see the generation of power or electricity, they partly depend on the wind variations or solar variations in these configurations. On the other hand, the cooling [demand] also partly depends on the outdoor temperature. 

So now think about the data center build out. So do we want to build data centers in regions with water stress or [in a] region where the stress will get worse in the future? Or do we want to build the data center at a better place? or there are also other considerations for potential risks such as the flooding and hurricanes. So by working together with climate researchers, climate [practitioners], the operation of the data center can potentially become more efficient and also the infrastructure will be more resilient to potential weather and climate problems.

[00:21:15]
Jeff Frame: [And] that's a great point. Nobody's typing in ChatGPT and asking it, are we going to have a warm or cold winter next year? That's not what we're doing here in terms of using AI for these predictions.

[00:21:28]
Gan Zhang: Yes, that's right. So I know some of the data centers, they have been dynamically allocating the load of their jobs. So if one region has let's say weather related stress problem such as a heat wave, they might allocate some of the jobs in other regions, so they are actually using weather and climate information to [guide] their operation.

[00:21:54]
Steve Nesbitt: So that's all we have time for today. Gan, thank you for joining us.

[00:21:57]
Gan Zhang: My pleasure.

[00:22:01]
Steve Nesbitt: Gan [Zhang] is an assistant professor in the Department of Climate, Meteorology and Atmospheric Sciences at the University of Illinois [Urbana-]Champaign. Jeff, thanks for joining us on this segment of Weather [Realness]. I'm Steve Nesbitt and

[00:22:11]
Jeff Frame: I'm Jeff Frame, also a professor in the Department of Climate, Meteorology and Atmospheric Sciences at the University of Illinois.

[00:22:20]
Steve Nesbitt: Have a weather or climate question you want us to tackle? Leave us a voicemail at 217-333-2141 or email weatherrealness@illinois.edu. Coming up next, we're breaking down the environmental impact of AI.