Look at a digital map of the world with pixels that are more than 50 miles on a side and you’ll see a hazy picture: whole cities swallowed up into a single dot; Vancouver Island and the Great Lakes just one pixel wide. You won’t see farmer’s fields, or patches of forest, or clouds. Yet this is the view that many climate models have of our planet when trying to see centuries into the future, because that’s all the detail that computers can handle. Turn up the resolution knob and even massive supercomputers grind to a slow crawl. “You’d just be waiting for the results for way too long; years probably,” says Michael Pritchard, a next-generation climate modeler at the University of California, Irvine. “And no one else would get to use the supercomputer.”
The problem isn’t just academic: It means we have a blurry view of the future. It is hard to know if, importantly, a warmer world will bring more low-lying clouds that shield Earth from the sun, cooling the planet, or fewer of them, warming it up. For this reason and more, the roughly 20 models run for the last assessment of the Intergovernmental Panel on Climate Change (IPCC) disagree with each other profoundly: Double the carbon dioxide in the atmosphere and one model says we’ll see a 1.5 degree Celsius bump; another says it will be 4.5 degrees C. “It’s super annoying,” Pritchard says. That factor of three is huge — it could make all the difference to people living on flooding coastlines or trying to grow crops in semi-arid lands.
Pritchard and a small group of other climate modelers are now trying to address the problem by improving models with artificial intelligence. (Pritchard and his colleagues affectionately call their AI system the “Cloud Brain.”) Not only is AI smart; it’s efficient. And that, for climate modelers, might make all the difference.
The ‘Cloud Brain’ might make it possible to tighten up the uncertainties of how the climate will respond to rising carbon dioxide.
Computer hardware has gotten exponentially faster and smarter — today’s supercomputers handle about a billion billion operations per second, compared to a thousand billion in the 1990s. Meanwhile a parallel revolution is going on in computer coding. For decades, computer scientists and sci-fi writers have been dreaming about artificial intelligence: computer programs that can learn and behave like real people. Starting around 2010, computer scientists took a huge leap forward with a technique called machine learning, specifically “deep learning,” which mimics the complex network of neurons in the human brain.
Traditional computer programming is great for tasks that follow rules: if x, then y. But it struggles with more intuitive tasks for which we don’t really have a rule book, like translating languages, understanding the nuances of speech, or describing what’s in an image. This is where machine learning excels. The idea is old, but two recent developments finally made it practical — faster computers, and a vast amount of data for machines to learn from. The internet is now flooded with pre-translated text and user-labelled photographs that are perfect for training a machine-learning program.
Companies like Microsoft and Google jumped on deep learning starting in the early 2010s, and have used it in recent years to power everything from voice recognition on smart phones to image searches on the internet. Scientists have started to pick up these techniques too. Medical researchers have used it to find patterns in datasets of proteins and molecules to guess which ones might make good drug candidates, for example. And now deep learning is starting to stretch into climate science and environmental projects.
Microsoft’s AI for Earth project, for example, is throwing serious money at dozens of ventures that do everything from making homes “smarter” in their use of energy for heating and cooling, to making better maps for precision conservation efforts. A team at the National Energy Research Scientific Computing Center in Berkeley is using deep learning to analyze the vast reams of simulated climate data being produced by climate models, drawing lines around features like cyclones the way a human weather forecaster might do. Claire Monteleoni at the University of Colorado, Boulder, is using AI to help decide which climate models are better than others at certain tasks, so their results can be weighed more heavily.
But what Pritchard and a handful of others are doing is more fundamental: inserting machine learning code right into the heart of climate models themselves, so they can capture tiny details in a way that is hundreds of times more efficient than traditional computer programming. For now they’re focused on clouds — hence the name “Cloud Brain” — though the technique can be used on other small-scale phenomena. That means it might be possible to tighten up the uncertainties of how the climate will respond to rising carbon dioxide, giving us a clearer picture of how clouds might shift and how temperatures and rainfall might vary — and how lives will likely to be affected from one small place to the next.
So far these attempts to hammer deep learning code into climate models are in the early stages, and it’s unclear if they’ll revolutionize model-making or fall flat.
The problem that the Cloud Brain tackles is a mismatch between what climate scientists understand and what computers can model — particularly with regard to clouds, which play a huge role in determining temperature.
Typical global climate models have pixel sizes far too large to see individual clouds or storm fronts.
While some aspects of cloud behavior are still hard to capture with algorithms, researchers generally know the physics of how water evaporates, condenses, forms droplets, and rains out. They’ve written down the equations that describe all that, and can run small-scale, short-term models that show clouds evolving over short time periods with grid boxes just a few miles wide. Such models can be used to see if clouds will grow wispier, letting in more sunlight, or cool the ground by shielding the sun. But try to stick that much detail into a global-scale, long-term climate model, and it will go about a million times slower. The general rule of thumb, says Chris Bretherton at the University of Washington, is if you want to cut your grid box dimensions in half, the computation will take 10 times as long. “It’s not easy to make a model much more detailed,” he says.
The supercomputers that crunch these models cost somewhere in the realm of $100 million to build, says David Randall, a Colorado State University climate modeler; a month’s-worth of time on such a machine could cost millions. Those fees don’t actually show up in an invoice for any given researcher; they’re paid by institutions, governments, and grants. But the financial investment means there’s real competition for computer time. For this reason, typical global climate models like the ones used thus far in IPCC reports have pixel sizes tens of miles wide — far too large to see individual clouds or even storm fronts.
The trick that Pritchard and others are attempting is to train deep learning systems with data from short-term runs of fine-scale cloud models. This lets the AI basically develop an intuitive sense for how clouds work. That AI can then be jimmied into a bigger-pixel global climate model, to shove more realistic cloud behavior into something that’s cheap and fast enough to run.
Pritchard and his two colleagues trained their Cloud Brain on high-resolution cloud model results, and then tested it to see if it would produce the same simulated climates as the slower, high-resolution model. It did, even getting details like extreme rainfalls right, while running about 20 times faster.
The ‘Cloud Brain’ tends to get confused when given scenarios outside its training, such as a much warmer world.
Others — including Bretherton, a former colleague of Pritchard’s, and Paul O’Gorman, a climate researcher at MIT, are doing similar work. The details of the strategies vary, but the general idea — using machine learning to create a more-efficient programming hack to emulate clouds on a small scale — is the same. The approach could likewise be used to help large global models incorporate other fine features, like miles-wide eddies in the ocean that bedevil ocean current models, and the features of mountain ranges that create rain shadows.
The scientists face some major hurdles. The fact that machine learning works almost intuitively, rather than following a rulebook, makes these programs computationally efficient. But it also means that mankind’s hard-won understanding about the physics of gravitational forces, temperature gradients, and everything else, gets set aside. That’s philosophically hard to swallow for many scientists, and also means that the resulting model might not be very flexible: Train an AI system on oceanic climates and stick it over the Himalayas and it might give nonsense results. O’Gorman’s results hint that his AI can adapt to cooler climates but not warmer ones. And Cloud Brain tends to get confused when given scenarios outside its training, such as a much warmer world. “The model just blows up,” says Pritchard. “It’s a little delicate right now.” Another disconcerting issue with deep learning is that it’s not transparent about why it’s doing what it’s doing, or why it comes to the results that it does. “Basically it’s a black box; you push a bunch of numbers in one end and a bunch of numbers come out the other end,” says Philip Rasch, chief climate scientist at the Pacific Northwest National Laboratory. “You don’t know why it’s producing the answers it’s producing.”
“In the end, we want to predict something that no one has observed,” says Caltech’s Tapio Schneider. “This is hard for deep learning.” For all these reasons, Schneider and his team are taking a different approach. He is sticking to physics-based models, and using a simpler variant of machine learning to help tune the models. He also plans to use real data about temperature, precipitation, and more as a training dataset. “That’s more limited information than model data,” he says. “But hopefully we get something that’s more predictive of reality when the climate changes.” Schneider’s well-funded effort, called the Climate Machine, was announced this summer but hasn’t yet been built. No one yet knows how the strategy will pan out.
The utility of these models for predicting the future climate is the biggest uncertainty. “That’s the elephant in the room,” says Pritchard, who remains optimistic that he can do it, but accepts that we’ll simply have to wait and see. Randall, who is watching the developments with interest from the sidelines, is also hopeful. “We’re not there yet,” he says, “but I believe it will be very useful.”
Climate scientist Drew Schindell of Duke University, who isn’t working with machine learning himself, agrees. “The difficulty with all of these things is we don’t know that the physics that’s important to short-term climate are the same processes important to long-term climate change,” he says. Train an AI system on short-term data, in other words, and it might not get the long-term forecast right. “Nevertheless,” he adds, “it’s a good effort, and a good thing to do. It’s almost certain it will allow us to improve coarse-grid models.”
In all these efforts, deep learning might be a solution for areas of the climate picture for which we don’t understand the physics. No one has yet devised equations for how microbes in the ocean feed into the carbon cycle and in turn impact climate change, notes Pritchard. So, since there isn’t a rulebook, AI could be the most promising way forward. “If you humbly admit it’s beyond the scope of our physics, then deep learning becomes really attractive,” Pritchard says.
Bretherton makes the bullish prediction that in about three years a major climate-modeling center will incorporate machine learning. If his forecast prevails, global-scale models will be capable of paying better attention to fine details — including the clouds overhead. And that would mean a far clearer picture of our future climate.