Paul Rand: Extreme weather events like hurricanes, heat waves, and once in a century floods are becoming more common and unpredictable.

Tape: Across the west tonight, a scorching March heat wave making winter feel more like the height of summer. From California to New Mexico, 38 million people under heat alerts.

Tape: This morning, streets looking like canals with neighborhoods underwater and people trapped after Broward County, Florida was swamped with one third of its annual rainfall in just seven hours. 

Pedram Hassanzadeh: Of course, the general public really care about extreme events, both in terms of early running, short-term forecasting, as well as the important question of how did climate change.

Paul Rand: That’s Pedram Hassanzadeh, associate professor of geophysical sciences at UChicago. His research examines how to predict extreme weather events and our changing climate using computational and AI models. 

Pedram Hassanzadeh: So I was always interested in this question of how to better simulate this system and get these extreme events correctly.

Paul Rand: For decades, weather forecasting has relied on massive models running on powerful supercomputers, but traditional models and AI sometimes struggle to forecast extreme weather events if they’ve never seen them before.

Pedram Hassanzadeh: So that’s a limitation of the current AI models. They do a lot of things amazingly well, but when it comes to things that they haven’t seen before, they fail.

Paul Rand: Scientists call these super rare events gray swans, and they are now trying to train models to observe patterns that might just be able to predict them.

Pedram Hassanzadeh: Now, one thing I’ve learned about these AI models is they’re not magical, but they always do better than expected. So they don’t do magic. They don’t extrapolate. They don’t predict things they haven’t seen. But one interesting thing we saw, and this was what very unexpected is they actually can learn from other regions in the world. So if they have seen a category five hurricane over the Atlantic Ocean, but they have never seen them over the Pacific Ocean, they can forecast them over the Pacific Ocean.

Paul Rand: These AI models could be trained to handle even more data, which could predict rare events. And Pedro and his colleagues saw success when models were able to predict intense monsoon rains in India during its peak agricultural season.

Pedram Hassanzadeh: And in real time, we generated forecasts that the Indian government sent via text messages to 40 million farmers last summer. The use of AI for weather forecasting will just keep increasing. To what degree the AI would reshape climate modeling, long-term climate modeling, I think that’s something to be seen.

Paul Rand: From the University of Chicago Podcast Network, welcome to Big Brains, where we explore groundbreaking ideas and discoveries that are changing our world. I’m your host, Paul Rand. Join me as we meet the minds behind the breakthroughs. Today on Big Brains, how AI is transforming the science of predicting extreme weather events. 

Paul Rand: Pedram, you are studying extreme weather, whether it’s hurricanes or heat waves or floods. Why is it so important for science to study this type of stuff right now?

Pedram Hassanzadeh: That’s a great question. So extreme events are, I think, important and interesting, at least for two reasons. One is they have major societal impact. They have a very disproportional impact on human health, on the infrastructure, on the ecosystem. You can have a heat wave just for a week, and it can lead to thousands of deaths. It can cause significant financial damage. Hurricanes, for example, can have, again, lead to displacement of tens of thousands of people or more, a lot of major socioeconomic impact. So of course, the general public really care about extreme events, both in terms of early warning, short-term forecasting, as well as the important question of how with climate change, these events are going to change. Also, I think just from a scientific perspective, they are interesting because they are very difficult to understand, model, and predict. And we know that these kind of things are just harder to understand compared to typical weather, but also just in terms of modeling, they are much more difficult to simulate.

One, property of extreme events is that they are rare. They’re infrequent. So just to get enough samples from computers or from observations, or if you do lab experiment, just to have enough of them to understand the underlying physics, to understand the statistical properties is just very difficult.

Paul Rand: Can you tell me where the science currently stands and how do you guys look at this and figure out whether a specific event was a natural variability or something that has been made worse because of climate change?

Pedram Hassanzadeh: Yeah. So that’s a major question that I think drives a lot of the research. Certainly my research since 2014, I can say that heat waves, it’s very likely that they’re going to get worse with climate change. We just know that the earth is warming, right? So there is a shift in the distribution of the temperature. We are going to get worse heat waves, also add to that humidity. So moist heat waves, which are very deadly, are going to increase. Now, if you ask about hurricanes, I think that’s still like a matter of research, how the frequency of the hurricanes change, how the size of the hurricanes, how just the surge, there are many components, right? But answering these questions are not easy exactly for the reason that you mentioned. The earth system has natural variability, right? 

Paul Rand: Right. 

Pedram Hassanzadeh: It has natural variability because changes in the sun, other changes change in the orbits over millions of years.

There is a natural variability there, but also as a part of that natural variability, there is something called internal variability. It’s a high dimensional non-linear system, so atmosphere, ocean, land, ice, they all interact. So they just generate oscillations. And El Nino is like one example of internal variability. We know it’s very important. There are many of these modes that basically change weather and the climate at different time scales. Now, when we get a heat wave and it is very intense, maybe record breaking, it becomes difficult to answer the question of was this driven by climate change or was this part of the natural variability? And currently the best we can do is that we can try to quantify how much climate change, the anthropogenic climate change might have increased the likelihood of this kind of events.

Paul Rand: Okay. So the way historically it’s been done is you take the current state of the atmosphere, you plug in your physics equation, you hook it up to the fastest supercomputer you can get at the time and let it run. And my understanding is from the time you’ve talked about is that forecasting has basically improved about one day per decade in terms of how far out we look. Is that pretty accurate? And that’s why our ability to forecast out longer has been indeed happening as computing power strengthens.

Pedram Hassanzadeh: Yes. And there’s a very nice paper led by Peter Bauer. It’s a paper in nature that’s called the quiet revolution of weather forecasting from like 1950s to late 2000s that they kind of showed this a steady improvement in the forecasted skill. It’s a result of two things. One, increasing the computing power so we can just resolve higher, like smaller and smaller scales in the atmosphere and also improvement in the observational systems, with better satellites, basically a better observational network. So we just can estimate a better initial condition for the physics-based model. So these two together have resulted in this quiet revolution, what people call the first revolution in better forecasting that was done by combining observations and computational simulation.

Paul Rand: Okay. And so is it that we’ve actually both hit a wall in this approach? And I mean, I guess computing will hopefully continue to grow and superpowers get stronger, but I hear that we’re somewhat of a wall.

Pedram Hassanzadeh: The major problem in basically dealing with the atmosphere and the climate system in general is that it’s very multi scale. So of course, we care about, okay, what is the temperature of a Chicago at the maybe length scale of tens of kilometers, right? We don’t need to know exactly like right outside my office at this moment how this parcel of air is moving, but the issue is that all these smaller scales, centimeters, meters, they all affect that larger scale that we care about. So you really cannot ignore these small scales. So what has been the problem in like when you say we hit evolve, the problem is currently like the best weather models do the simulations at the global resolution of 10 kilometers. So to solve everything larger than 10 kilometers, below 10 kilometers, you need to come up with some functions that tell you, for example, how the smaller scales down again to centimeters and millimeters fit back into 10 kilometers and larger.

And these functions are very ad hoc. There has been a lot of research on them, but still we are not ... And there are various limitations in our understanding and computational how we can develop these codes. This has been the problem. Same problem with climate predictions. There are dysfunctions, they are semi-empirical, there are a lot of knobs in them. So when you see, for example, uncertainty in the track of a hurricane, a lot of it comes with this, what we call subcutiscale parameterization, these functions that whatever NOAA uses could be different from what the European center uses, which could be different from the function that another group uses. So we get a range of forecast because of these differences. So this is really at the heart of the problem.

Paul Rand: So what we all hear, whether we’re listening to a forecast or reading about it, is this idea of a climate model. And I wonder maybe for us to understand what in the world a climate model is, is it like a giant simulation or what do those models actually look like and what are the limitations of the current models that you’re working with?

Pedram Hassanzadeh: So climate models are basically, they’re built on weather models. So they basically solve these equations called never stocks equations that govern the conservation of mass momentum energy of any fleet flow, whether it’s the blood in our veins or whether it’s some flow in the atmosphere or in some galaxy far away, it’s the same equation. It’s a very company- Far far away,

Paul Rand: Not just far away.

Pedram Hassanzadeh: Very far away. It’s the same equation, but then how to solve those equations on these supercomputers has been a challenge. And again, this problem of multi-scale. So there was a student at UChicago, student of RSP who really developed these equations that are more suited for climate predictions, not to do weather forecast for a few days, but longer term. And that’s the Princeton group, again, implemented this and did the first climate modeling that Suki Manabe, who received a Nobel Prize in physics a few years ago. He was the guy who was at Princeton, really started this transition from, okay, we cannot do weather forecasting, let’s go to climate predictions, which are again, these set of equations, written encodes, millions of lines that really solve the equations for the motion of air and water and how the energy and mass and everything moves in the atmosphere and ocean.

And then we basically do predictions of that.

Paul Rand: Okay. Now what you’re talking about, maybe the best thing is we’ll call that the first revolution, but you talk about AI as the second revolution in weather forecasting. And you started looking into this from what I’ve read back in about 2017, and that’s pretty early on. And so I’m wondering what made you think that AI could do something that the super computers weren’t at that time able to do? Yeah.

Pedram Hassanzadeh: So I was always interested in extreme better events. So I was always interested in this question of how to better simulate this system and get these extreme events correctly. And at some point, I became very interested in what you can directly learn from data rather than having all those empirical functions that I mentioned before. Can we just directly learn from data? And then I was an assistant professor at Rice University and I had this student who told me, “Hey, I have done some work with these neural networks.” And that was really the beginning, 2017, that there were a few groups around the world who were thinking about, well, can we use these deep neural nets and the artificial intelligence to do something about weather and climate? And so we were trying to, just for fun, we thought, okay, let’s just try some of these. And I actually, after we did some work, we realized, okay, this idea of what we call analog forecasting, learning from weather patterns has been around for a long time.

And we saw that actually the neural nets can do a great job in learning from these patterns. And back then, I mean, most people were telling us this will not work because you don’t have enough data to train this giant neural nets, but there were a few groups that were doing this. And so my students and I, we ended up collaborating with a group led by Anima Anakamundar, who is a professor at Caltech. Back then she was also leading a big group at Nvidia. And we developed this model called Forecast Net that it turned out you can just have a giant neural net, you can input the current state of the atmosphere, it can predict the next one. And you can just train this on ... We have around 40 years of very high quality data from 1979 when satellites went up and did really comprehensive measurements of the atmosphere continuously.

And we saw that you can train on these 40 years and you can predict the weather almost with the accuracy of the best physics-based model, except that the neural network is much, much, much cheaper, like around 100,000 times faster, 10,000 times more energy efficient. You can basically do a state-of-the-art weather forecast on a laptop rather than requiring a supercomputer.

Paul Rand: Okay. So basically the idea of what’s happening now is that AI is allowing you to run a lot more simulations, if I’m understanding this right, than you would’ve been feasible before. How does running more simulations actually improve our confidence in predicting extreme weather?

Pedram Hassanzadeh: So as I mentioned, one problem with the extremes is that they are rare, right? And so you basically have a low signal to noise ratio, right? So having more simulations, more samples that you start all of them from a little perturbation in the atmosphere, because we don’t have exact estimate of the state of the atmosphere, right? So the butterfly effect is that even if you have a very tiny difference in what is the temperature you think the atmosphere has and what is the temperature, you are going to, in a few days, get different forecasts. Now just imagine if I can just do many perturbations to the state, run the neural net. So just to give you an idea, like the European Center, which is the leading weather forecasting center in the world, they can afford around 100 what’s called the ensemble members. So when there is a hurricane coming, they can do 100 forecasts and that would give them a cone of uncertainty.

And as I mentioned, also the different weather models are different because they use different of the functions for their smaller scales, right? So then you also have to have five models, run each of them 100 times. This is extremely expensive. But now with the neural nets, you can have five of these, you can easily generate thousands of these forecasts. And so you can really try to get a better sense of the uncertainty of the extremes. This is one major reason. The other one is they also directly learn from data and it turned out for many things, like especially precipitation. Just directly learning from data is better than having this physics space model with some of those ad hoc subdued scale parameterizations.

Paul Rand: Okay. So check me if I understand this correctly, because my understanding is that what AI is allowing us to do is push reliable weather forecasts from about eight days out to about nine or 10. It doesn’t sound like a lot, but it is a lot. Help me understand why.

Pedram Hassanzadeh: Yeah, that’s a good point. So you can gain a day maybe in the forecast scale, but again, this comes at a much, much lower cost.

Paul Rand: Okay, so cost is a big part of this.

Pedram Hassanzadeh: If you just think about the electricity bill of a weather forecasting center, and also the infrastructure that needs. So I should say that when we did this work, combination of the work that was done in the AI community, in the weather climate community, a lot of it was led in the tech companies. This is where the excitement came from that, hey, we have these models that are more accurate than what we had before. And one day you mentioned could be like a result of 10 years of model development with highly skilled people around the world work very hard to do this, but it’s faster, you can do more ensembles, but there is more into it. These models are public, they are open source, they’re easy to use. And so when I moved to University of Chicago, I gave a talk in a launch that Epicat as part of the Climate Institute, and I mentioned Forecast Net, some of this work.

I thought I’d giving a talk to some students who are like interested in having lunch while I’m speaking. And then this person came to me and he said, “I’m very interested in this work. My name is Michael Kramer.” And it turned out, so he’s a mobile prize winning economist at U Chicago who has been very interested in weather, building some of the work he was doing on weather. And since that day, this is the hardest I’ve ever worked in my life because it turned out

For an economist like Michael to see that, hey, we can now really use these AI models to escalate the better access around the world because one thing we keep forgetting is that sure, European center, US, they have great weather models, they have the supercomputers, they have the highly skilled scientists and engineers. Many people around the world don’t have that. So people in Africa, people in many other low middle income countries, they cannot develop their own models. They cannot generate their own forecast. So they cannot take a model and improve it, tailor it for their own region, right? With AI models, all of this can be done very easily. So this is, I think, an aspect of this revolution that was not really seen by many, including me in our communities until we learned about this work that Michael and Amira and others have been doing.

And that led to what we call the human center weather forecasting at University of Chicago, part of the Climate Institute is really let’s use this AI revolution, but make the weather forecast accessible and really tailor to people all around the

Paul Rand:

World.That’s terrific. I’m imagining that there are also some patterns that AI can find that humans or traditional models may have historically missed. Can you talk about that a little bit?

Pedram Hassanzadeh: Yes. So there are, of course, I mentioned we have like 40 years of high quality observations, which is really not that long. And the earth, just the natural variability has modes that the time at scale is like decades, hundreds of years, even longer. We have some observations maybe from 1850, but they become very low quality and maybe AI can help with filling some of the gaps. So that’s what some people are working on. But with AI, there’s this big question of, well, okay, how about events that if we just had a longer observation of earth, we would have seen, but we haven’t seen before. And these events we call gray swans, events that are physically possible, but they haven’t happened before. So one idea is that if I have a good AI model and I let it run for a very long time, can it actually generate some of these gray swans that we just haven’t seen before?

Paul Rand: What do you call these gray swans?

Pedram Hassanzadeh: Gray swans are events that are very physically possible. We just haven’t observed the system enough to see them.

Pedram Hassanzadeh:

And these are the events we don’t want to miss. These are really the events we care about.

Paul Rand:

Got it. And so you tested this with your forecast net, right? Yeah. That you did it with NVIDIA?

Pedram Hassanzadeh: We thought, okay, gray swans by definition, we haven’t seen them, right? So it’s not like I can test the model and a gray swan, but let’s do it the other way around. So we have the 40 years of data, let’s just remove very strong extremes from the training set. And so what we did, we actually took out any day that there was a category three, four, five hurricane anywhere around the world. So we trained ForecastNet on this new data set. So this is a model that has seen almost four years of data, except that it has not seen those days that had a category three, four, five hurricane.

Paul Rand: I see. Okay.

Pedram Hassanzadeh: Interesting. So then the question is, can this model forecast a category five hurricane? Because it’s a gray swan. We know it happens, right? It’s physically very possible, but can we test the model? And it turned out it cannot. So that’s the limitation of the current AI models. They do a lot of things amazingly well, but when it comes to things that they haven’t seen before, they fail. Now, one thing I’ve learned about these AI models is they’re not magical, but they always do better than expected. So they don’t do magic, they don’t extrapolate, they don’t predict things they haven’t seen. But one interesting thing we saw, and this was what very unexpected is they actually can learn from other regions in the world. So if they have seen a category five hurricane over the Atlantic ocean, but they have never seen them over the Pacific Ocean, they can forecast them over the Pacific Ocean.

Paul Rand: But it has to learn it from somewhere else first.

Pedram Hassanzadeh: Exactly. But I think that makes them very powerful because it sense that, I mean, there are many places in the world that it rains a lot, right? So then if we are going to get some extreme rain event in one region that had never seen it, and these are where actually you get the biggest impact of the extreme events when it happens somewhere, people are not used to it, then the Omada already has learnt it from another region. So something we did that to show this is, if you remember around a year and a half ago, well, two years ago, there was this flooding event in Dubai because of this unprecedented precipitation, like rainfall over Dubai in April, which was twice more rain than any rain in that region ever. So we thought, okay, can the AI models forecast this because they have never seen this kind of rain over there.

And it turned out they actually can do a very good job. Because we figured out they have seen them in other regions in the world. I mean, Dubai is a dry region. So again, they’re not magical, but they can do a lot of things already. So now the question is, can we better understand it and help them to even get these gray swans and the extreme events?

Paul Rand: Okay. As you talk about teaching the models, one of the concerns, criticisms that often comes up about AI is that a lot of these models are black boxes and they give you an answer, but you can’t see how they got there. What does that mean in the context of weather forecasting and does that matter or we just want the output?

Pedram Hassanzadeh: Yeah, I think it matters a lot. I mean, weather forecasting is an important thing, right? You don’t want to have the wrong forecast, tell people in a city like Houston to evacuate and then nothing happen. Or you don’t want to miss like an extreme event. And of course, we can do our best to benchmark these models and get a sense of the performance, but especially for extreme events, because they’re very rare, you don’t have enough samples to test them. Gray Swan, you even have never seen them. So how would you know how the model would do for a gray swan? So understanding how these models work and when these models fail, it’s I think extremely important. And I think it’s a very exciting research question. So, but back to your question about black box, the problem is, like for example, forecastnet is a model with 75 million learnable parameters.

So it has several layers. There are a lot of neurons that talk to each other. When you want to train it, there’s 75 million numbers that you need to learn. And the newer models get that to around a billion. Wow. So billion numbers they learn and then they can do better forecast for you. The issue is, I cannot look at one of these neurons and tell you, this is the one that forecast the hurricane and this is the one ... So when the model fails, this is the neuron I need to go and fix. So this is the problem. And so I think as scientists really want to have a better understanding of this so that we can go and fix that neuron. Or another idea that’s emerging is that it’s really not one neuron doing something, but it’s a combination of a couple of these neurons working together, which I think is also a question generally about AI, about LLMs.

I think that there’s a lot of great work going on in the board of LLMs about how the different parts work together and then end up with models that maybe lie. So you want to be able to go and constrain that part of the model that’s responsible for be dishonest. We really want to be able to do the same video weather models, to be able to figure out what part of the model is doing what, so that I can go and fix it.

Paul Rand: Got it. In terms of progress and speed, is there a point where you look at this and say that you’re going to have a confidence that AI is going to allow us to eventually replace some of these more traditional climate models or arguably even the meteorologists themselves, as you hear about in so many different fields?

Pedram Hassanzadeh: Yeah. So I think for weather, forecasting up to like two weeks, I think now there are AI models that are operational, but probably when you look at your phone, part of the forecast would come from an AI model. I’m not a fan of replacing all the physics-based weather models with AI models. I think that would be dangerous. Again, if a gray swan shows up, I think a physics-based model has a better chance of forecasting it, but I think complementing the existing weather models with the AI models because of all those advantages in the speed and things like that, it’s already happening. It’s like the operational weather centers are very excited about this. The world metrological organization now has all the working groups. I’m part of one trying to help countries around the world to use AI. I don’t think this will replace the metrologist in any sense.

I think at the end of the day, we need metrologists who know the region who can look at the output of these models, make a decision. I think what will change is that of course now every metrologist needs to have an understanding of how these models work, when they fail, when they are great. So it’s just changing the tool box, right? But I think human needs to be in the loop, expert need to be in the loop for this. Now, this is about weather. Climate, which is like, when we talk about climate models, it’s about predicting, for example, how the likelihood of temperature above something over Chicago would change 20 years from now, 30 years from now, depending on how much CO2 we are going to put in the atmosphere. That problem has been much more difficult because we don’t know the future, right? So if I bring a neural net, I cannot show you that my model is accurate, different from weather, it’s more like extrapolation problem, it’s out of sample.

So it’s much higher bar to think about how to use neural nets AI models for climate prediction. So that’s, again, another area of research. There are several ways that we think this can be done, but I think that revolution has not been as fast as the weather forecast revolution.

Paul Rand: Okay. Let’s talk about real world impact and how this is all coming. And we made the comment about, well, is a day or two a big deal in predicting, but in actuality, a day or two could be a life saving big deal that could really save many, many, many lives. Talk to me about that, but I also wonder on a bigger picture, how does this research help communities prepare, but not just with forecast, but how they build infrastructure and plan for emergencies?

Pedram Hassanzadeh: Yeah. Let me give you an example of work that we have done. So after I met Michael Kramer and with Amir, we started this human centered weather forecasting. What we did was, we looked at how these models can, like a bunch of these AI weather models can forecast the onset of the monsoon in India. We know that the monsoon in India is a very important part of the life of over a billion people, especially that affects the agriculture. Michael and Amir had worked in the past with the Indian government to provide some forecast based on some statistical models, but the problem was scaling it up so that it can provide accurate forecast better than what the farmers currently have access to all over India. So we benchmarked some of these AI models. We found that actually a couple of them have very good forecast skill for the monsoon onset.

And we blended them, we closely worked with the Indian government. This was actually a project of the India government. And in real time, we generated forecast that the Indian government sent via text messages to 40 million farmers last summer. And it turned out that this was an interesting year because monsoonis started early in South of India and usually when the monsoon hits that region called Crawler, then farmer That’s the different parts of India. Okay, if it hit Crawler today in two weeks or in 18 days, it arrives in my region. Got it. This year, that didn’t work because Mansoul started and then it just stopped. It just didn’t move for two and a half weeks. But the farmers received our forecast, which told them that Monsoon is coming later than expected, even though it was all over the news that this is a year with earlier.

Oh my gosh. So this actually, and Michael and Amir are now doing work that they talk to the farmers to see who used the forecast, did they benefit from it? They really quantify- That’s a

Paul Rand: Great story.

Pedram Hassanzadeh: Yeah. The measure. But this was an example of these AI models being very useful. Now, could we have done it with the traditional models? They also can get some of these, but AI models are so easy. We at U Chicago generated the forecast. We did it with a bunch of these forecasts. We can now go and even further modify the forecast. Next year, build a model that would be better for Mansoon onset. We are now currently replicating all of this over Ethiopia, working with the Ethiopian Metrological Institute. So there’s a lot of advantages in AI that comes from the speed, being open source, being public, being easy to change. That again, with the traditional models has been very difficult, but of course the AI models are also more accurate. So the question is, can we develop better tools to better inform the public, to better inform the government, the local governments about how to do adaptation, how to prepare as the climate is changing?

Paul Rand: Take me out 10 years. And we talked about what we can or can’t predict if we haven’t seen it yet and so forth. 10 years time, how do you think we may be talking about this?

Pedram Hassanzadeh: Well, I think the use of AI for better forecasting will just keep increasing. Okay. To what degree the AI would reshape climate modeling, long-term climate modeling? I think that’s something to be seen. I’m actually writing a perspective paper about this. So I started thinking more deeply. I think we have to be more careful because again, we don’t have the data from the future. We cannot verify the quality of these models, but I think there will be a big impact from AI for also climate prediction. There is a third question to what degree AI might help us discovering new things about climate physics, things that traditionally have been done with people sitting in this department and doing math and experiment and looking at data. That is the part that I think is most exciting, but also I think it’s the part that has been the slowest part.

And to what degree some LLMs can work together, look at data and come up with a new hypothesis about how the atmosphere works. I think that’s something that it’s hard to predict. We already hear this in some fields like materials, science, biology. In climate, it hasn’t got there. We haven’t seen anything new being discovered, but I think that’s also another very exciting direction that I’m very interested in do this kind of thing. Already there is evidence that of course these LLMs, for example, combined with other things, can accelerate the research in climate, but again, to what degree they can really change how we do the research and let us discover new things about the earth system. I think that’s something you will see.