As temperatures in recent years broke historical records, Zeke Hausfather, a climate scientist with the research nonprofit Berkeley Earth, tried to find words to describe the heat, settling on “absolutely gobsmackingly bananas.” He also sought new ways to visualize it. Hausfather created his own graphics, including a striking “tree ring plot,” with the help of a brainstorming and coding assistant: ChatGPT.
“It’s interesting, in part, because it’s not what I would’ve expected AI to be good at,” Hausfather said.
Like millions of other people, climate scientists are finding a role for large language models in coding, communication and other parts of their workflow. They’re also pointing AI tools at central questions: How hot will it get, how rainy, how fast?
“AI is offering some pretty exciting opportunities to tackle questions we’ve been stuck on for a while,” said Elizabeth Barnes, a Boston University professor who specializes in environmental data science. “But it is not a complete transformation of our science.”
That’s because traditional and AI climate research tools are likely to complement each other. Climate models are complex programs that simulate the physics of the Earth system with equations requiring more than a million lines of code. Scientists refer to them as physics-based models to distinguish them from AI and other models, which work without simulating physics. Climate models can project large-scale change that hasn’t happened yet, but they have trouble resolving influential small-scale phenomena like cloud formation.
AI tools may be able to infer values for those things, but they can’t yet “see” outside their training data — for example, to extreme weather occurring on a scale outside the historical record. (There’s initial evidence that they can transplant extreme weather from one part of the world to a region where it’s never occurred.)
Scientists are publishing new approaches to AI every month, across a broad spectrum of topics. Here are a few themes:
Global risks go local
Insurers, homebuyers and lots of other people want hyper-localized estimates of climate risk — like a property’s odds of flooding — in order to make confident financial decisions. The trouble is that global simulations are far too coarse to be of use on such a small scale. Researchers are hopeful that AI can help punch big-picture simulations down to local levels better, by combining model results with historical weather data so that the AI “learns” how they’ve been related in the past.
Hybrid AI and physics-based modeling may provide a “flexible, accurate and efficient way” to attack the problem, Google scientists wrote last spring.
Google Research this week released Groundsource, a tool for predicting flash floods, which cause more deaths than any other water-related hazard. Researchers used Gemini, the company’s large language model, to identify 5 million news articles since 2000 chronicling 2.6 million flash floods in 150 countries. Using this data and machine-learning-trained weather models, the team produced a publicly available tool (not yet peer-reviewed) that yields valid results 82% of the time.
Welcome, ‘co-scientists’
The Atlantic Meridional Overturning Circulation (AMOC) carries warm water north and cool water south. The current is a stable feature of the global climate that, among other things, keeps Europe warmer than it otherwise would be. It also faces a risk of collapse over the next century, which scientists are eager to better understand.
Google DeepMind led a team of more than a dozen scientists, including Hausfather, that last month released a preprint assessment of the state of the AMOC. They undertook the work specifically to test how well an AI “co-scientist,” Gemini, could collaborate on a broad review of current science, akin to comprehensive UN assessments.
The team synthesized 79 papers about the AMOC and revised their work 104 times over 46 total person-hours — which is roughly 10 times faster than it usually takes. Almost all of the material the AI contributed to the review was kept, and it made up 42% of the final version.
Climate experts have knowledge and intuition that isn’t a part of AI training, though. While the AI is a useful collaborator, it’s nothing close to a stand-in. “Substantial oversight was required to expand and elevate the content to rigorous scientific standards,” the research team wrote of the work by their Gemini helper.
Cloudy projections
How clouds affect the flow of heat in and out of the atmosphere has long beguiled scientists. Low-lying clouds bounce sunlight back to space; high-altitude clouds trap heat below. How they form and how they are changing are difficult questions that strongly influence climate projections.
Combining AI with physics-based estimates is showing promise in side-stepping the hardship of trying to simulate clouds. A team of university, nonprofit and corporate scientists in 2024 concluded that machine-learning techniques “could be used as a replacement” for current cloud estimates in some models.
Regional climate change
Scientists sometimes find discrepancies between global model projections and what’s occurring on the ground locally. Both physical models and hybrid AI models have trouble reproducing, for example, the drying out of the US Southwest. They tend to see rising temperatures bringing more moisture to the area, not less, as the global atmosphere swells with water vapor.
That’s a hard problem for AI, too, but it’s come closer to real trends than other methods.
Tiffany Shaw, a University of Chicago atmospheric scientist, and colleagues are conducting a series of studies to benchmark what AI models can do today, so they can track progress with improvements. That includes a current working paper in which they show AI outperforming the models and hybrid models in the drying Southwestern US, “although we still don’t fully understand if [AI is] getting it for the right reasons,” Shaw said.
University of Washington scientists in August published a study showing that an AI model trained only to minimize errors in short-term weather forecasts ended up being able to simulate observed cyclone activity in the western North Pacific Ocean.
Elsewhere, AI is turning raw satellite data into readings of methane emissions, simulating how glaciers calve and identifying weather extremes long before they happen.
It’s clear that AI tools, like the LLMs, can perform many tasks as well as or better than people. But even when they produce the right answers, scientists still need to grasp how they arrived at them to move the field forward.
“Science is about not just the outputs, but the understanding that goes behind them,” Hausfather said. “One of the challenges with AI is that it’s hard to understand what it did unless you can dig into it.”
This report is auto-generated from Bloomberg news service. ThePrint holds no responsibility for its content.
Also Read : New book champions AI-Climate nexus. It’s co-written by Amitabh Kant