AI weather and climate prediction face real-world tests
Artificial intelligence-based computer models are starting to augment standard, physics-based tools used to predict the weather and climate change. In the next few weeks, they face real-world tests with high stakes.
Driving the news: With a snowstorm targeting the Northeast this weekend — followed next week by an intense Midwest blizzard, then a major U.S. Arctic outbreak — computer model skill will be a major topic in the minds of weather and climate scientists. Many of them are not yet sold on the AI model's accuracy.
Why it matters: The adoption of AI in meteorology and climate science is likely to be one of the most game-changing developments for these fields since the dawn of numerical modeling.
The latest: Newly released AI-driven computer models like Google DeepMind's GraphCast will be put to the test during the next few weeks as winter weather becomes far more active across North America and Europe.
- This will be the public's first opportunity to see how such models handle significant and often difficult-to-forecast winter storms, compared with projections that come from more widely known tools like the GFS and European model.
- At the same time as weather applications move forward, companies are working to advance AI approaches to making reliable climate projections amid claims that the planet's warming even faster.
Zoom in: The AI-driven models are trained on historical data to learn complex systems and generate projections, whereas numerical models use physics equations — and weather observations — to produce simulations of future conditions.
- Companies that are getting into weather and climate AI include Google, whose GraphCast model has proved especially accurate compared with the elite models forecasters currently rely on.
- Others include Nvidia, IBM, Tomorrow.io, along with government agencies like NASA and NOAA.
Yes, but: There is considerable skepticism in the weather and climate community about relying too heavily on these models, given their novelty, and the fact that forecasting has steadily grown more accurate through the use of physics-based numerical models that run on supercomputers.
- Also AI models suffer from perception issues. Whereas forecasters generally understand the physics equations that go into conventional models, AI-driven ones can seem like projections being spat out from a black box.
Between the lines: The AI models like Google's GraphCast are far faster and cheaper to run, able to be completed within minutes compared with several hours for conventional models.
- But AI's results are only as good as the data they were trained on. This suggests that it may be possible that extreme weather events from climate change, many of which are unprecedented, could be missed entirely.
- However, so far that hasn't been the case, according to Hendrik Hamann, IBM's lead scientist for climate and sustainability research.
- "The way I would explain it to people who have sort of their home in numerical weather prediction models, is to say look, everything we are doing is we are creating a representation of the physics based on observations, or maybe reanalysis but observations, ideally, eventually."
- "And that physics is as true as it is in a numerical prediction model."
What they're saying: In an interview with Axios on the sidelines of the COP28 climate summit in Dubai in December, Hamann said AI models are not as mysterious as they may first appear.
- "We are certainly going in the direction that AI will play a much more important role. It won't replace simulation-based numerical weather prediction models, that will take a while, but it will definitely start more and more to impact that," he said.
- "So the resistance, I [am] honestly saying, I'm from physics, too. I kind of can appreciate it," Hamann said. "But I think it is based on assumptions about AI which may not be true."
- Hamann noted that numerical weather prediction models also make many assumptions, and in some senses are a black box too.
The intrigue: Hamann said that foundation models, which IBM is developing, including via a partnership with NASA, are particularly useful for exploring AI-driven approaches to weather and climate science.
- Foundation models are trained using massive weather and climate datasets and can be used for different applications without being extensively modified.
- For now, teams building such models are relying on so-called "reanalysis" data generated by computer models, which show historical weather conditions. This makes them models learning from the output of other models.
- Hamann called foundation models "the digital twins of weather" and said the goal is to ultimately pipe weather observations into them, rather than reanalysis results.