It’s a complex problem that we face with the climate – and it would take a very powerful computer to be able to really model the Earth System and its future with any granularity.
I recently listened to Mike Pritchard, a research manager at Nvidia and professor at UC Irvine, talk about this process and how it works.
“The physics span 10 orders of magnitude in space in time,” he said, citing research problems like figuring out whether a cloud particle favors water vapor congregating around it.
“If you want to simulate the planet hundreds of times, to sample many ‘what if’ scenarios of the future, you unfortunately, even with the most powerful super computers, can’t do justice to all of that complexity,” he said. “Meanwhile, humanity’s questions about the future climate are too broad for simulation technology.”
For a concrete example of looking at problems directly, he talked about his commute from San Diego to Irvine, and seeing a particular kind of cloud out of the window.
“It looks like a gray band on the horizon,” he said. “We call it the marine layer. If it wafts in on a beach day, you’re bummed out because it makes you cold. But what matters is that it’s the edge of a massive sheet of low clouds that you’ll see out your window from the airplane, half the way on the flight from San Diego to Hawaii, and that cloud reflects a lot of energy from the planet, keeping it cooler than it would be otherwise. So if it dissipates … that will amplify global warming … but if it thickens up, which it could, that will damp it. And that’s a multi-trillion-dollar uncertainty. And it’s a simulation problem. We know these clouds take very high resolution to simulate that we can’t afford to deploy in climate simulation yet.”
A Cavalcade of Systems
Pritchard also mentions the word ‘ensemble,’ which is often used in machine learning to talk about utilizing more than one model at a time, or crowdsourcing the outputs of different LLMs, but it has a different meaning in weather prediction.
“You don’t predict one hurricane,” he said, “you predict hundreds of hurricanes. You hope for the best, and plan for the worst … card-carrying atmosphere scientists at the University of Washington are taking these AI weather models, which were trained on the mess of the real atmosphere, which is very noisy and messy, and then after the fact, probing them and asking if they’ve learned physics by doing things like this.”
Pitchard talked about how that works with technology, and building an archive of evidence for the ability of AI models to help us with weather prediction.
New Software and Technology
As an example, Pritchard mentions the capability of Nvidia AI tools such as Modulus, Earth2Studio that enable the research, development and validation of such AI Forecast Models.
The company, which has vaulted to the top of the pack on the U.S. stock market, actually has many research streams and collaborations with the atmospheric science community. These get released in the open source domain and here are some of the prominent models :
StormCast research – demonstrates a generative AI model that emulates atmospheric dynamics, looking at mesoscale weather phenomena and making predictions (paper).
CorrDiff– this is another generative AI model that creates high-resolution weather forecasts (paper). One can learn and explore more about AI powered downscaling using pretrained Corrdiff here.
FourCastNet– this model achieves 25-km -resolution weather forecasting for places around the world with Spherical Fourier Neural Operators, recently calibrated for huge ensembles. One can learn and explore more about medium range global forecasting using pretrained Fourcastnet here.
Earth-2 Platform is a digital twin cloud platform that is helping enterprises leverage these AI advances and accelerate traditional numerical simulations to reduce the computational bottleneck of climate and weather simulations. Coupling these advances with advances in computer graphics like RTX rendering technology, we can build digital twins of earth’s climate and weather to help scientists explore, analyze and explain the complexities of weather phenomena especially in the context of changing climate.
More on Climate Work
Pritchard also talked about promoting optimal dispersion in large ensemble AI weather predictions, and referenced new papers that bring more detail to the emerging science of using AI to simulate low-likelihood, high-impact climate extremes. This, he said, will give climate risk modelers new tools to help us understand and guard against extreme weather events.
Backward and Forward
Here’s another aspect of what Pritchard talked about in terms of useful AI models. He described traditional climate informatics processes as like ‘going to an Oracle’ – large simulators create large data sets, he suggested, which users then have to mine to help inform ‘what if’ scenarios and questions about future climate. AI predictions, he added, can run forwards and backwards, which should help users more easily figure out what could have changed, given a different initial input.
“We might be entering a future where we can understand our influence on the future more easily, without having to experience all the bottlenecks of conventional simulation,” he said.
The Power of Twins
In conclusion, Pritchard also talked about the idea of digital twinning applied to the biggest single item that we have in our world – the world itself.
“I think that the really important paradigms of interactivity are chains and cascades of AI digital twins, … So you can imagine a future that’s evolving towards AI digital twins of the climate, coupled to AI digital twins of extreme weather (events).”
Giving the nod to current research and what everyone is doing around this very complex problem, Pritchard gives us food for thought on how to address the climate of our time with technology that goes far beyond big data sets. Stay tuned for more of what came out of recent events on AI and the planet here in Boston.
