How AI could shape the future of climate science

While “artificial intelligence” might conjure thoughts of chatbots or streaming recommendations, AI has also made a name for itself as a weather forecasting tool — prompting climate scientists to ponder whether machine learning and other types of AI algorithms could bring about a similar revolution in their field.

“We saw the power of these algorithms, but we also saw the challenge of being able to roll them out for a long time because they accumulate error the same way a simulation can accumulate error,” explained Laure Zanna, professor of mathematics at New York University, about the challenges of using AI to study a system as complex and dynamic as our climate.

Now, recent advances in the stability of machine learning models have allowed climate researchers to more thoroughly explore AI as a tool. During sessions organized by the APS Topical Group on the Physics of Climate at the Global Physics Summit, scientists shared achievements and discussed the future of research at the intersection of AI and climate science.

While AI might be new to this field, physics has long been an essential part of understanding Earth’s weather and climate, said Francesco Ragone, assistant professor of applied mathematics at the University of Leicester, who highlighted this contribution during a tutorial at the summit. “There's a long history that has intersections with fields of physics which are not necessarily directly related to climate science — like statistical physics or dynamical systems theory — that are the backbone of our understanding of the climate system.”

Building on this grounding in physics, what can AI do to make existing climate models better? One of AI’s main advantages is speed, as a model for a system as complex as the climate requires a lot of detailed equations. “Some of the equations [in climate models] are known, and you can model them, but you don't have infinite computing [power],” said Zanna.

One strategy to tackle this complexity is to develop hybrid models that have a core physical model but can learn new patterns from data with the help of machine learning. This hybrid approach lets researchers take advantage of “data and tools that we didn't have decades ago,” said Tapio Schneider, an environmental science and engineering professor at the California Institute of Technology. “I think the biggest opportunity [with hybrid models] is to drive the physics forward [by] learning from data,” he added.

During the summit’s two “AI applications in Weather and Climate” sessions, many talks focused on using AI to learn from climate data, said Ching-Yao Lai, an assistant professor of geophysics at Stanford University and session co-chair. “We still rely heavily on existing physics-based models,” said Lai. “[But now], there are emergent examples of leveraging all the observations we have, then directly learning new physics from those observations.”

AI can also help with model parameterizations — the process of replacing small-scale or complex processes with simplified versions — by allowing researchers to “integrate large amounts of data with preexisting physics-based models [to parameterize] processes that are still poorly understood,” said Lai. One example, presented by Schneider, was CliMA, a hybrid physics-AI model that runs on the cloud and can incorporate up to 100 terabytes of data.

There’s also been progress in using AI for climate emulators, or models designed to mimic a specific system by “machine learning the whole dataset,” said Zanna, who presented Samudra, an emulator that can model physical interactions of the ocean, atmosphere, and sea ice.

Along with sessions on statistical physics, dynamical systems, and more traditional climate physics research, the AI-focused climate talks at the summit covered a broad spectrum of topics, including advances in machine learning algorithms, methods for predicting sub-seasonal weather patterns, and climate modeling with quantum computing algorithms.

Researchers have more work to do to realize AI’s full potential in climate science. “The challenge for AI/machine learning models is still to predict climate change — the new dynamics and patterns, with more extreme events, that the models have not seen before,” said Jörg Schumacher, professor of mechanical engineering at Technische Universität Ilmenau and session co-chair, in an email.

But it's a challenge that scientists are motivated to tackle, as this research could lead to faster, more precise climate models that use less computing power and enable the “discovery of physics we have not seen before,” said Lai.

Schneider is hopeful that advances in AI could vastly increase both the precision and speed of existing climate models. “We've seen something like a 10% improvement per decade in key metrics,” he said. “I want to see 50% improvement — so five decades worth of progress realized in a much shorter time.”

Zanna believes that AI can continue to bolster the field if researchers continue to use a multi-pronged approach. “It's [going to be] a combination of physics, data, and simulations that are going to be one of the breakthroughs,” she said.

Ragone added that while AI’s future impacts look promising, he’s also hopeful to see continued research on more precisely quantifying uncertainties in models, which can lead to better predictions.

“The presence of [GPC] this year at the [Global Physics Summit] has been the largest to date — there has always been a niche group of people working on these things, but in the past few years, with the recognition of the Nobel Prize and of course with the emergence of AI applications, I think this [field] is trending upward quite a lot,” said Ragone.

Global Physics Summit registrants can access on-demand presentations until June 30.