Last week, I served on a panel at a National Academies of Sciences Roundtable on Artificial Intelligence and Climate Change workshop titled Accelerating Climate Progress with AI: From Science to Action Workshop. The workshop was an incredible gathering of government scientists and officials, professors, industry experts, nonprofit leaders, all focusing on the same topic: how do we use AI for action on climate change?
I’ll get to this question later in the blog, but a better question to start with is: how and why exactly do climate scientists use AI in their work to begin with? This blog is the third and final blog in a three-part series on climate science. The first two blogs were on the history of climate models and what a climate model is.
Artificial Intelligence: another tool in climate scientists’ toolbox
Climate scientists use many tools to conduct research in understanding Earth’s complex climate system and how it might respond to an increase in heat-trapping pollutants. Of course, the main tools we most often read about are climate models, which are extremely complex approximations of the physical climate system. While they technically date back to the 1960s, climate models have been improved and made more complex over decades, attempting to mimic or model the climate system (atmosphere, oceans, ice, land surface, and vegetation).
AI is merely another tool in our toolbox. Specifically, most climate scientists use machine learning, a type of AI that, in principle, is similar but very different in practice from Generative AI (GenAI), which can make predictions by learning patterns from large amounts of data.
For example, you might be familiar with El Niño, an increase in ocean surface temperatures in the East Pacific that occurs every two to seven years. When an El Niño event occurs, it affects weather around the world often in predictable ways. For example, the US Pacific Northwest is usually warmer during an El Niño event, while the Southern US is typically cooler and wetter than usual. And this is why AI is such an incredible tool for climate scientists: our climate system is full of patterns in data.
AI can easily identify how El Niño impacts the weather around the world by learning that a warmer East Pacific Ocean usually results in a warmer US Pacific Northwest. And this is just one example of a pattern in our climate system: once AI is given a large amount of climate data, it can learn statistical patterns in how the climate system has behaved and use those patterns to make predictions.
It is worth noting that climate models and AI in climate science are vastly different tools: a climate model is based on a set of equations that describe the physical properties of the climate system, while AI learns statistical relationships. This is important. AI is just learning patterns of the climate system to make its predictions without explicitly following the physics of the climate system, so its predictions are technically un-physical.
Artificial intelligence in climate science research
Already, AI is being leveraged quite extensively in climate science and has actually been used for nearly 25 years (far before the GenAI boom). One exciting example is something called a climate emulator, the first of which was developed by non-profit Ai2. Climate emulation is an AI approach that quite literally emulates a climate model: it learns all the patterns of how the atmosphere responds to changes in ocean temperature so that it can make its own predictions. The Ai2 Climate Emulator (ACE) is already being used in climate research, since it can be run 100x faster than a traditional climate model and is nearly 1000x more energy efficient. Plus, it can robustly replicate the predictions of a real climate model!
Another example of leveraging AI in climate research is something called downscaling. In my climate model blog, I mention that climate models simulate the Earth as a giant three-dimensional grid and calculate how variables (like temperature) change at each grid point. Unfortunately, these grid points are spread out so much so that it’s impossible to analyze how temperature or precipitation might change on a local scale.
This is where AI comes in: it can make predictions about how the climate changes on a local scale by learning patterns in climate model data and localized observations. So instead of saying broad conclusions like, “the Northeastern US will warm due to climate change,” downscaling with AI could provide specific values for Boston or New York City.
These are just two examples of how climate scientists leverage AI in their research. There are plenty of other examples, including combining AI with climate models to create a hybrid climate model. It is also worth noting that the AI used in climate research is much less energy intensive than the large language models like ChatGPT. There isn’t a particular need for data centers to power the kind of AI research climate scientists do.
Artificial intelligence in my research at UCS
As climate change produces more extreme weather and climate disasters around the world, more international focus is being placed on climate litigation, or suing fossil fuel companies for climate damages, and the operationalization of the United Nations Loss and Damage Fund. Innovations in attribution science, a sub-field of climate science that attempts to answer the question, “how much more likely was this extreme event due to climate change?” will help further advance climate litigation and the L&D fund.
For climate attribution science to be robust, scientists need a long and reliable record of climate data. Unfortunately, many locations in the Global South lack these critical data records, making attribution science studies less reliable in these regions, despite the fact that they are the most likely to be affected by climate change.
In my research, I use machine learning to fill in data gaps in Global South historical records of extreme temperature. How do I do this? We know that AI can make predictions based on what patterns it learns from the climate data it’s given.
In my case, I give the machine learning model a substantial amount of daily temperature data from multiple high-resolution climate models, making it learn common temperature patterns during extreme heat events, and then ask it to make a prediction of what the temperature would be at the locations that are missing data. We test how well the AI model does by comparing it to other observational datasets. Filling in data gaps with machine learning is an exciting avenue of research.
Highlights from the workshop and issues with AI
AI in climate science has so much potential. As I’ve written above, academics, nonprofits, and industry are already using AI in climate research. But how exactly could AI help with bridging climate science and climate action?
At the National Academies of Sciences workshop last week, we heard some incredible examples of how AI is already being used for climate mitigation and adaptation purposes. Some outstanding examples include using AI to: predict the final size of a wildfire from ignition, detect wildfires early, advise sustainable farming practices, advise farmers on how they can adapt their crop to climate change, provide recommendations for tree planting in urban environments, and achieve sustainability for major water aquifers throughout the country.
AI isn’t all rosy though; there are plenty of issues with this new tool. First, it’s difficult to use AI to make some of the aforementioned predictions of a future climate since AI cannot learn patterns on data from the future (it doesn’t exist yet!). In other words, projecting climate change with AI is not an easy task.
AI is also a sort of black box: we don’t know what AI is doing to make its predictions. Explainable AI is a tool that is starting to be employed more by climate scientists. It can help reveal where and what kind of patterns AI is looking at in climate data to make its predictions.
And of course, we as scientists need to do more to build trust in AI. There needs to be a clear history of rigorous evaluation of AI in science; there need to be published, open access, and reproducible AI research methods; and there needs to be explicit documentation of how confident we are in the data the AI models learn patterns from.
AI is here and it’s likely to stay. Climate scientists will continue using it for research and for action, but again, it’s just another tool in our toolbox. It won’t be taking over the field entirely anytime soon.