SC20’s keynote was delivered by renowned meteorologist and climatologist Bjorn Stevens, a director at the Max Planck Institute for Meteorology since 2008 and a professor at the University of Hamburg. 

In his keynote, Stevens traced the history of climate science from its earliest days through the modern-day brink of the exascale era. He likened the evolution of computational climatology to Deep Thought, the gargantuan supercomputer from Douglas Adams’ novel The Hitchhiker’s Guide to the Galaxy. In the book, scientists task the supercomputer with delivering the ultimate answer to life, the universe and everything. 

“Deep Thought … thinks deeply for 75 million years, and it comes up with the answer to the ultimate question that they posed, but by that time the descendants of the people who built the computer forgot the question,” Stevens recounted. “So they went about and built another computer, one out of organic components, and they called it Earth – and the idea then was to go into the computer and learn more about the question that they got the ultimate answer to.”

“And so if you keep that in mind, you’re all set for today’s story.”

Meanwhile, 130 years ago…

To explain his analogy, Stevens took the audience back to 1890. At that point, he said, Svante Arrhenius was developing a seminal equation to explain the surface temperature of the Earth.

“So he worked out this idea that you see here: that … the average surface temperature of the Earth could be related to how radiant energy (that’s ‘R’) … flows through the system,” Stevens said. “And he thought about a bunch of other things, too – so he introduces the idea of ‘H’ – that’s horizontal heat transport – might matter, and he introduces ‘V,’ the idea that the vertical heat transport might matter, ‘O’ (other things) could matter, and ‘C’ was the carbon dioxide and water vapor – the greenhouse gases.”

So: if one understood radiant energy (R), horizontal heat transport (H), vertical heat transport (V), greenhouse gases (C) and other things (O), one could explain the Earth’s surface temperature – and crucially, they could predict how fluxes in those variables would change it.

“The problem was, [Arrhenius] didn’t really understand any of them,” Stevens said.

“The problem was, [he] didn’t really understand any of them.”

Stevens ran through the various meteorologists and climatologists who worked on solving the variables in Arrhenius’ equation: Vilhelm Bjerknes did some early work on horizontal heat transport; John von Neumann used the ENIAC computer to flesh horizontal transport out even more; Fritz Möller worked on radiant energy; Joseph Smagorinsky put together a lab to build circulation models of Earth.

Enter Syukuro Manabe. Circa 1967, Manabe worked with both Möller and Smagorinsky to solidify radiant energy and approximate vertical heat transport. Using computer modeling, they developed the first compelling representation of how the surface temperature could change if you increased greenhouse gases – but, Stevens cautioned, there were lots of exclusions and approximations.

Over the years, many other things (O) got added to the equations and models. By 1979, with H and R more or less solved for, and greenhouse gases (C) standing as the independent variable in which researchers were most interested, there remained one glaring stumbling block.

V for vexation

V: vertical heat transport.

“The big thing is getting the V right,” Stevens said. “And as pretty as people make these sorts of models look, they don’t help in the least with V.”

But why does this one variable pose such a huge problem relative to the others? As it turns out, it has to do with the thinness of Earth’s atmosphere.

“The thinness of the atmosphere is on one hand a really wonderful thing, because it means that [horizontal heat transport] involves a sort of quasi-two-dimensional circulation,” Stevens explained. “There’s not a lot of vertical going on. Transporting energy from the equator to the pole happens in a thin atmosphere, and you don’t need to worry so much about the vertical to get it right… but the thinness becomes a problem when you think about V.”

Stevens pointed out that the Earth’s circumference is around 40,000 kilometers; from the equator to the pole is around 10,000 kilometers; and the eddies that move energy horizontally through the atmosphere span around 1000 kilometers. So, he said, you can get away with resolving to around 1000 kilometers (“they could do that in the seventies,” he added).

“But if you wanna get V,” he said, “you need two orders of magnitude. You don’t need a thousand kilometers: you need ten kilometers, because the atmosphere’s so thin.” 

This factor of 100, however, appears deceptively achievable. Factoring in the additional temporal and spatial dimensions of climate modeling, the computational requirements of calculating V relative to those of calculating H skyrocket from an increase of around 2^7… to an increase of around 2^28.

“So it’s a massively larger calculation to get V,” Stevens said, “and in 1979 you had absolutely no right to expect to get it.”

“It’s a massively larger calculation … and in 1979 you had absolutely no right to expect to get it.”

That’s Moore like it

Leaning on Moore’s Law, Stevens mused on the expectation that chip performance would double roughly every 18 months. “But really, what kind of technology has exponential growth for what seems like forever?” he asked. 

“Well, it turns out it’s your kind of technology,” he answered, presenting a graph of how computing power has boomed over the last 50-odd years. “Having exponential growth that goes on forever and ever and ever like this… is just crazy.”

“So you could ask yourself: well, in retrospect we had half a century of doubling. … Maybe V isn’t so impossible after all. What do we need?”

Stevens did the back-of-the-envelope math. 28 doublings (remember 2^28?); 18 months per doubling. 42 years, he concluded with a smile – though he stopped short of explicitly acknowledging the synchronicity with Hitchhiker’s Guide.

So: 42 years between when researchers reliably solved for H and when they could reliably solve for V.

1979 plus 42 years.

2021.

“Practically tomorrow,” Stevens said.

The future is now…

Tempering excitement, Stevens cautioned that while many of the tools are in place, the climate science community isn’t quite there yet – and much work remains to be done. Storms and clouds still vex researchers, and may (or may not) impose additional performance requirements beyond the expected needs. 

Climate models, he explained, require a certain resolution to be useful – roughly 100 simulated days per day at a grid resolution of roughly 1.5 kilometers. Citing his team’s current results, he said they would only need around 100 times more computing power to achieve acceptable numbers.

“That’s already there,” he said. “That’s in place. I mean, if you look at the JUWELS booster, we could do this calculation that we want, we could get the throughput that we want, but maybe we could only do it once.” Elsewhere, he added, researchers had already started to achieve the necessary resolution at regional levels. 

“Machines like Fugaku and the emergence of exascale [are] showing that not only are these calculations feasible, but they’re going to be practical.”

“Machines like Fugaku and the emergence of exascale [are] showing that not only are these calculations feasible, but they’re going to be practical.”

… but what next? And why?

Returning to his allegory, Stevens said that achieving this resolution in the exascale era was comparable to receiving that first answer from Hitchhiker’s Guide’s Deep Thought: an answer so long in the making that the question may be long forgotten by the time it is delivered. So, he explained, we needed to take that next step, asking ourselves what the questions are and ensuring that our results become useful in the world.

To do that, he suggested, we would need to diminish the barrier between computation and interaction. “The old way of interacting with machines involves experts and layers and layers of expertise,” he said. He brought up a picture of a couple of children playing with a tablet. “Here we see the future,” he said. “As brilliant as these kids probably are, they don’t know CUDA. They don’t even know Python! They probably don’t even know English! … But they’re interacting with a machine.”

This, he said, was the endgame: to take these supercomputers and “expose their information content to users who can work through the consequences of their actions, of their policies, of their imaginations, of their hypotheses.” By way of example, Stevens cited the European Union’s Green Deal, which incorporates a plan for massive Earth simulations called “digital twins” that will be run on supercomputers and which are designed to allow researchers and decision-makers to explore the results of scenarios and policies.

“The Atlantic… It looks like Van Gogh.”

Throughout most of the keynote, Stevens’ tone was that of an energetic, geeky professor: excited about the future, fascinated by the research in his field, happy to make a sci-fi reference. Toward the end of the presentation, though, his tone noticeably shifted as he brought up a satellite image of the Atlantic Ocean in mid-September of this year.

“The Atlantic… It looks like Van Gogh – the Starry Night picture – because you see swirl after swirl after swirl,” he said. “These are hurricanes, tropical storms, and pre-tropical storms, and tropical depressions – there’s eight of them, count them – eight of these storms!”

“You might see one storm, kind of cool, you look at its eye… you might see two if you’re really lucky. Once in a great while you see three. And here we have eight! Climate change? Who knows.”

A pause.

“Wouldn’t you like to know?”

“Wouldn’t want to know what we’re doing to our Earth if it’s causing things like this to happen?”