SHF2026｜Bjorn Stevens

Bjorn Stevens

Director and Scientific Member at the Max Planck Institute for Meteorology;

Fellow of the American Geophysical Union (AGU)

It's really a pleasure to share this stage with the very stimulating presenters that we've heard so far and that we'll hear shortly. Special guests, dignitaries, colleagues, friends, I'd like to talk about what I consider one of these Anthropocene problems. If you think about what we've been discussing today, we are talking about a class of problems that have the character of scaling globally, whereas governance seems to scale locally. If you look at nuclear proliferation, finance, climate, and AI, these are all technologies or human products that have a global impact, but which we struggle to regulate or govern because the capacity of governance we have tends to be more local. Climate is a sort of canonical Anthropocene problem. I'm going to talk to you a little about some new tools we have for managing the new reality of climate change. The new reality reflects the fact that we are moving into a new era.

For most of my life, we anticipated changes in the climate system that we could explain. The world would warm as we increase the amount of CO₂. We're seeing those changes now, but this new era also means that we're seeing changes that we didn't anticipate and which we cannot explain.And thankfully, we have some new tools which we hope will help us surmount these difficulties. This is an example of one which I'll elaborate on a bit further later.

In this figure, what you see is a period over 26 years where the black line is steadily increasing. This black line is due to measurements by satellites like the two that you see on the upper right-hand part of that figure, two NASA satellites.It's been a program of measurements for 26 years looking at how energy is flowing into the Earth system and flowing out. And the black line documents the increase in energy flowing into the Earth system. It's been increasing at about a rate of a half a watt per decade.

This increase is what we would expect from increasing greenhouse gases. So that, in some way, is no surprise. But this plot was made a few weeks ago as we gathered 40 experts from around the world in a small castle in the southern Bavarian Alps, because the rate of increase is something that we can't explain.It's increasing about twice as fast as we would expect. And if we take everything that we know, the models that we use, our understanding of how CO₂ interacts with radiation, what we end up with is an estimate that the planet is three or four times more sensitive to CO₂ than we thought. We're not ready to make that conclusion scientifically, but what this black curve illustrates is that there are big things going on that we don't understand.

Either we can't measure things as well as we thought, the curve might be wrong. Things are happening that our theories and models don't incorporate, so the models might be wrong. Or maybe the planet is just more sensitive than we were led to believe.

The other really disturbing thing about this figure is that the instruments that are keeping track of this essential quantity are going away. These satellites are at the end of their life. There's no known program for replacing them.And there's very few organizations that are in a position to maintain vital measurements of the Earth, such as the ones that we had before and the ones that accompanied the old order that we heard about earlier today. So this is one issue where we're seeing new changes that we can't explain.

Another area where we're having difficulties is that we know the Earth is warming and it's warming because of CO₂, but as we look a bit further there's things that we realize that we'd like to know that we don't know.

This is a pattern showing 35 years of observed warming. Where you see red, it has been warming more than average, and where you see blue, it has been warming less than average. This pattern has caught us by surprise.

If you take our existing models and our existing theory, the pattern we expected looks more like this: much more warming in the Northern Hemisphere (red) and less in the Southern Hemisphere (blue), as opposed to what we actually observe, which is more of an east-west asymmetry. Differences in temperature patterns like this have profound consequences on the shape of weather patterns, where and how intense hurricanes or typhoons will form, how rains will fall, and how atmospheric rivers will cause flooding.

So being able to say something not just about the planet warming as a whole, but where it warms and how fast it warms, is something we're finding out we can't do as well as we would like. The models really weren't developed to solve that problem, and they certainly weren't developed to solve these problems.

So these are familiar figures that you might see on the daily news. You can collect them any year. Normally I update this slide year by year with certain. This year is disasters. The tornado in the upper right rushing through the Iowa landscape is particularly ironic as it ravages the wind turbines.So you can keep your eye on that. But you see the flooding in Valencia and Spain. You see fires that no one expected wreaking havoc in the Southwestern United States near Los Angeles.You see storm surges and deluges associated with hurricanes. And I don't think I need to tell anyone in this audience when you look at the past summers of the amount of rainfall in cities like Beijing, but also across China, the scope and the magnitude of extreme weather that's developing, which we're totally unprepared to predict. We can talk about it qualitatively, but not quantitatively.

We don't understand how hurricanes seem to be intensifying so rapidly. And the models that we've been using to understand climate change weren't even developed to consider things like this. In fact, if you take a standard climate model that we use in international reports, it's been explicitly designed to dampen extremes and regress to the mean.

So the basic principle of how they work is to take an event like this and try to make it go away. So we're certainly incapable with our existing tools to deal with things like this. So why is it so difficult? What's the problem? Why are our models inadequate? And the simple answer is really shown in this wonderful figure from the Artemis crew.

I don't know if you appreciated the irony of this figure, but the canonical image in the world imagination was taken 54 years ago in 1972 from the Apollo 17 astronauts. And they took a picture of the earth that's known as the blue marble. And it looks like this, except Africa is the other way.So here again, we see the world and you struggle to figure out where you are in this world, but the big brown spot is Saharan in Africa. And if you look at the top of the figure, you'll see the South Pole. So it's like the Apollo figure, but turned a bit on its head.

And the interesting thing about this figure is it illustrates really how the earth is large and the atmosphere is thin. And that's the key challenge in terms of representing the earth system with more fidelity. That green aurora that you see forms through the ionization of oxygen and it forms in a layer at about 114 or 115 kilometers in the atmosphere.It just tells you how thin the atmosphere is.

And when you think above us, most of the weather that you would experience in a day of severe weather in Shanghai would occur in a much shallower layer, something about 10 or 20 kilometers deep. So in order to represent how atmospheric circulations transport water and influence cloudiness and form rain, you have to be able to simulate a very thin layer over a very large planet.

And that creates an enormous computational challenge. The degrees of freedom of a model that tries to represent that range of scale number about 10 to the 12. It's still about 1,000 times larger than the largest large language models that exist.And not only is it 1,000 times larger, but those degrees of freedom need to be evolved in time at fractions of a second so we can look forward to see how they evolve over years and centuries. But computers are big. And we can today simulate on this scale.

So this is an example of a simulation. It shows winds in colors and it shows clouds in white. And it recently was at the same GTC that our previous speaker was talking about, stood behind the stage, the simulation of Jensen Huang as he was presenting his new insights.

And it shows how we can simulate the Earth system now with a fidelity that we only previously could dream of. We could simulate down to the local scale here over Shanghai to get an entirely new look at how the Earth, the atmosphere, the turbulent ocean, the nutrient cycles, the chemical cycles respond to human activities. But we can take these models, we can simulate clouds and rain.The green is dynamic vegetation as it would evolve over China. In the middle figure we look at ocean currents in the East Pacific interacting with nutrient cycles which are brought up by the circulation. And on the bottom we look at the daily cycle as carbon fluxes in and out of the Southern Ocean.

Our ability to simulate the Earth system is profound given advances that we see in computing. And it offers the possibility of wholly new insights into how our world, our Earth, will respond to human activities. On the largest systems in Europe we can simulate this full Earth system at a rate of about 700 years per year.If you give me that computer for one year I can simulate 700 years in the future or 100 years seven times or seven years 100 times to explore varying scenarios of how the world might change to stimulate our thinking to explain some of the problems that I mentioned at the outset. The challenge of course though is that we can simulate the world with great fidelity. But if you look at the data that's produced, the output, it's overwhelming.

This one year of simulations of 700 years of simulated Earth translates into about 150 petabytes of data. It's okay, there's data centers that can host 150 petabytes, but try to read it. I'll make the model and I'll put the data on the server and try to do anything with it and you'll quickly realize you can't do anything. 

It's like this old story. There's a dilemma. It's called the cartographer's dilemma and a lot of people know it from Jorge Borges, the Argentinian author. It was first done by Lewis Carroll and hopefully you're reading along. But I'll summarize quickly for you. It's an Englishman talking to a German, it kind of fits.I live in Germany, Hamburg, the sister city to Shanghai. We're celebrating our 50th anniversary. But in this story the Englishman might be speaking to me, “mein Herr”, and they're talking about their maps.And the Englishman says, “how big is your map and how much detail does it have?” And the German says, “well, my map just has fantastic detail.” And they say, “well, how much? How many kilometers per inch do you have?” And at one point he says, “well, we have about six inches of map for every mile of the world.” And I said, “well, my map's even better.I can have, I have like six yards for every mile. And recently we made a map that has for every mile of the map, you have one mile of the earth. So you can see every detail, you can make it even bigger.”

But the problem is: how do you unfold this map? The mein Herr complains that they wanted to unfold it, but of course it covered the whole countryside, and the farmers complained. So they decided to just use the countryside itself as their map. That might work if you're navigating spatially, but for climate, we are navigating in time. 

So we don't want to wait until things happen to figure out where we're going. We have maps so large we can't unfold them. But what Lewis Carroll didn't anticipate was that AI gives a new dimension to humanity.We don't have the three dimensions of space and the one dimension of time. We have a fourth dimension, which is the silicon. We have the digital dimension, and we can unfold the maps in the digital dimension.We can send AI in and ask it to take the information we need out. And if you think about it, that's really what's happening with a lot of the AI we're seeing: its ability to extract useful information and provide it to us—information that is in principle there somewhere, but that you couldn't possibly figure out where to find in any reasonable amount of time. 

And so AI solves this paradox for us.It allows us to use these new physical modeling tools to understand how the earth works, and to use and become interactive with the information it creates. There are two exciting approaches. One is fairly straightforward emulation. You run the models, train the AI to emulate them, and then use the emulators to explore their information content. But there's another one which I find even more interesting: generative AI. This is what's often used to make pictures.And what generative AI does is it takes a training dataset—here, these worlds, 700 worlds or 700 years of simulation—and ingests them.

This is an example from NVIDIA, from a colleague, Noah Branovitz, who leads a team at NVIDIA. And what he showed was that you could take our model output, like what I showed you before, and learn the underlying distribution of the data.So the goal is to learn the probability distribution of this high-dimensional dataset. In his figurative idea, you put it in a bottle, and then you can go back and sample that distribution. The beauty of that is, if you're a practitioner planning a water system in Shanghai, developing a new agricultural system, working with a renewable energy network, or thinking about how to distribute crops around regions of the world, you can go and explore specific scenarios. You can say, given that dataset—the training data—what would be the biggest possible storm you could make over Shanghai? And then you can create that storm and draw it back from the data.

And here's an example: you go and say, give me a time of day, give me a year, give me a pattern of sea surface temperatures like the one I showed you at the very beginning, and say how would a typhoon look if it came right into Shanghai? What's the strongest one you could make? What's the second strongest one you can make? Then you can create phenomena that you can give to planners, where they can work out how best to plan to manage the future. You can give it to scientists to evaluate: does that really make sense? Is it really consistent with the data? You move from trustable AI to verifiable AI. On one hand, you enliven a scientific community to explore the realism of these effects, and you enliven a user community and an impact community to work out the consequences of these effects.And this is what I find particularly exciting: the use of generative AI as a way to represent the information content of models like the ones I showed.

An example is here. This is still quite primitive.Many of these things are still in development, but this is from another colleague at NVIDIA. They just show that if you take a map of the world, you can orient yourself. He says, let's pick a place in the world, given the training data that I learned from these physical models, make me a tropical storm over the Bay of Bengal somewhere.On a given day—in the red—we say, okay, it looks like we might have a tropical disturbance there, and you can see it developing a bit. It's fairly coarse resolution. Then you can go with the diffusion model and say, well, actually, let's look at this at super resolution.So you begin to see something that looks more like a tropical storm. You can say, well, I'm not really interested in the Bay of Bengal. Maybe I'm interested in the Arabian Sea.

I've actually always wondered what a typhoon would look like if you put it over the Strait of Hormuz, which would wrap around the mountains there and cause a deluge in cities like Dubai. How do you plan for that? Tools like this allow you to envision these scenarios and work with them. It reminds me of another story.We heard about Alibaba from the Arabian Nights. There's another one with Aladdin and his magic lamp. It says, in one of the large and rich cities of China, there lived a tailor named Mustafa. He was very poor. He could hardly, by his daily labor, maintain himself and his family, which consisted only of his wife and son. His son, who was called Aladdin, was a very careless and idle fellow, but he had the good fortune to come upon a magic lamp with a genie, and that brought him to great prosperity.

So I was wondering—we don't have too many students here today, but maybe there is an Aladdin who will use these tools to help inform society about how our planet is evolving and give us the tangible feel we need as a population to grow our minds in a way that we can conceive of things like global governance. Only with an ability to experience and see the worlds can we really hope that the people who put our leaders in power will develop the consciousness they need to choose wise leaders. So I like to think that maybe the sea bottle is an early version of Aladdin's lamp, and one of you out there will take it forward. And I think as you do, it creates the possibility of really creating new economies—new economies which are able to bring information about the future to life for people, and which aren't just fanciful creations of AI, but are faithful representations of scientific understanding.So here on the left, I show the scientific basis of fundamental models and observations which are being used to train AI. AI is only a representation of that data, but it's a representation that brings interactivity to that information for users from a variety of many different sectors. And that brings me to the end of my story.

I should say that the advantage of this is something that scales. In a way, this isn't different from what we've done before in very tailored ways. You would go hire McKinsey if you're rich enough, and you would have a consultant to help improve your waterworks. But why do that when you can do it this way, and you can scale that and make it valuable to every small agency that is planning their waterworks, urban planning, or food security? So that brings me to the end. I think we can see the outlines of how we can move forward to embed climate information in decision-making, or how to embed a global consciousness and a future consciousness in decision-making.

We are sailing uncharted seas, as we heard from earlier talks—not just in the climate realm, but in many realms. But we have this enormous technical capacity to help us as humans envision where we are going and perhaps make wiser choices about what to do about it. I say here the technical capacity is mostly there.

We have the physical models. We have ways to make them interactive through AI. But realizing this potential comes back to something we heard at the very beginning of the day.It comes back to questions of governance. We need access to the computing to make the virtual worlds that AI will then interpret. But we really need standards.We need ways—this comes back to the regulation of AI. If anyone can make an agent that can do anything, there's really no value in any of them. So what we need to do is establish standards that say: what is a faithful representation? What represents best practice? Only then, for instance, can companies actually deliver this information.

Imagine if you're a company building these agents and explaining to a country or a city how to develop their waterworks, and it completely fails. You might be liable. But if we have standards, we reduce the liability because you can show that you followed good practice. So without rules, we won't be able to have innovation. We won't be able to open markets. We won't be able to explore the potentialities that these technologies are offering.

And the other thing we need is a robust dialectic between research and practice. We need to bring these tools into practice in the public sphere and the private sphere, and we need the research and academic community to stay somewhat apart and in tension so that we can criticize the tools and engage a more productive development. And of course, we need continued and improved earth system modeling.

I think if we do this, this forms elements of an industrial policy that grows technical prowess in people, machines, and software. It can build new economies, and I think it would build soft power through global stewardship. So thank you very much for your attention.

(This article is edited based on the recording and has not been reviewed by the speaker.)