Earth scientists to environmentalists: AI isn’t all bad

If you ask environmentalists about artificial intelligence, they will likely say its biggest drawback involves energy use. One AI research company estimates that querying a service like ChatGPT or Google AI uses 30 times as much energy as a conventional Google search. Energy use is not the only environmental component of AI. Cooling the large data centers that house these AI tools requires enormous amounts of water, an increasingly scarce resource in many parts of the United States. One large data center can suck up the same amount of water per day as a small town. Others have pointed out that AI chatbots can spread misinformation about the climate crisis.

Many environmentalists have argued that AI should be significantly curtailed for these reasons. Perhaps unexpectedly, Earth scientists—researchers who study the environment to make life-saving discoveries—have found themselves at odds with this group of natural allies.

Whether it be for weather forecasting or earthquake detection, Earth scientists argue that AI—particularly machine learning, which is an AI technique that autonomously pulls insights from pools of data—is the most energy-efficient and fastest way to make new breakthroughs. “Traditional” weather forecasters and other researchers have exclusively used computers to solve equations that past scientists would have had to solve by hand. But even if a computer does these calculations much faster than a person, it still takes energy, and a lot of it. AI models for Earth science use pattern recognition and other inference tools to produce results, instead of explicitly solving equations, which brings down energy costs. But even for this application, AI, like any new technology, is not a perfect solution.

An efficient alternative. Despite energy concerns surrounding AI, the original appetite for AI use in Earth science was spurred by a need to manage computational resources.

To forecast weather, you need information like temperature, humidity, wind speed, and pressure. But you might need this information over hundreds or thousands of miles, potentially at different elevations. You might need hourly data for a period of days, or even weeks. The data volume piles up quickly, and running these variables through equations is costly in terms of both time and energy. And in the case of a traditional weather forecast, these operations don’t just run once. They run multiple times to create an ensemble of forecasts.

Each forecast is tuned slightly differently but uses the same equations to create future weather states. Weather isn’t deterministic. You can’t put variables into an equation once and know with complete certainty what the weather will be. Nature is too chaotic. An ensemble of forecasts is supposed to capture that uncertainty by providing the likeliest range of future weather. This is why, for example, your phone’s weather app reports the “chance of rain” as a percentage. For a physics-based weather ensemble to give an accurate prediction, it needs to contain about 50 individual forecasts.

Supercomputers that run these ensembles can’t service every project at once, however. Limited computing resources are shared with other researchers, so the queues to use these supercomputers grow. This could tremendously slow progress on both weather forecasting and climate research. By using AI models instead of—or in tandem with—physics-based models, these calculations can be done more efficiently, while still generating accurate results.

Instead of directly “calculating” potential weather (like rainfall rate or maximum daytime temperature), AI models trained on past data recognize relationships between input variables and potential weather states. For example, an AI model might predict pleasant, sunny weather today given yesterday’s temperature, humidity, and pressure, based on historical patterns.

The least sustainable part of running an AI weather or climate model is training the model. Training time and costs could be similar to those of running a physics-based weather ensemble. Once the model is trained, however, it never has to process another equation, making it orders of magnitude faster and more efficient. And the AI model might only need to be retrained on new data annually. That means thousands of forecasts can be issued at marginal cost and energy use before the model has to be trained again.

A better alarm system. Ignacio Lopez-Gomez, a research scientist at Google, and his colleagues on Google’s climate and sustainability team are developing a generative AI model that better predicts dangerous weather extremes.

Their model takes two physics-based forecasts and then emulates thousands of other forecasts based on them.

“We can issue thousands of potential forecasts for the future, and so that enables us to estimate the risk of low likelihood, but very high impact events,” Lopez-Gomez explains.

A 100-year flood event has a one percent chance of taking place. A traditional weather ensemble with 50 forecasts may miss that one-in-100 event. With thousands of forecasts the chances of identifying the possibility of those events skyrockets.

“Let’s say that they would have to close transportation methods, schools, it would be impactful for different communities,” he says. “It’s actually important to have that information, even though it’s very low likelihood.”

Another area in which AI can be useful is earthquake detection. Marine Denolle, professor of Earth and space sciences at the University of Washington, says seismology has a rich legacy of available data that makes it the perfect training ground for an AI model.

“Nuclear monitoring in the 60s fueled a ton of money into the operation and monitoring of earthquakes and nuclear explosions,” she says.

Over the last 60 years, seismologists have identified and labeled ripples in the Earth’s crust that are caused by everything from earthquakes to explosions. This provides the ideal dataset for training AI models to detect earthquakes. Having a model that can separate “contributions from different sources, like a car, or a thunderstorm, or an earthquake” can be more expedient than having experts manually differentiate these signals. This signal classification, which works by extracting features from the data, is a different technique than generative AI. As Lopez-Gomez says, generative AI is akin to “if you give a question, [the model] predicts the most likelihood answer.” In signal classification, the AI scans the material—images, waveforms, texts—and categorizes them based on the distribution of certain features. Cancer researchers are using such methods to predict how well patients might respond to certain treatments based on specific features of their tumors.

AI for earthquake detection is also a boon with a growing array of sensor technologies. Fiber optic lines and and smartphone accelerometers have added other ways to detect subterranean vibrations. AI models can sift through vibrations from these ad-hoc seismology tools to detect smaller earthquakes, locate their points of origin, and detect new faults.

“When the earthquake is small, its signal is going to be buried in the seismic noise, but the AI was trained to see through the noise, and so it can find these peaks of small events,” Denolle says.

Rapid identification of small tremors as earthquakes could provide additional life-saving notice to first-responders or emergency managers.

Energy trade-offs. Despite the high costs to train AI models, the energy use in training a weather or seismology model will likely never rival what large language models like ChatGPT require.

“In our space we’re still quite a few orders of magnitude away from what they are experiencing in large language models,” Lopez-Gomez says. “Large language models are trying to compact the entirety of human knowledge on the internet in a model.”

Because of the type of services large language models are trying to provide, they need to do trillions of operations. Weather and climate models, in contrast, have to do 10,000 times fewer operations. There’s just a lot less weather and climate data to train a model on, and it won’t grow in perpetuity. Better observations are increasing weather and climate data, but old data has to be phased out for model training. Climate change means that weather or climate states from 20 years ago may not be relevant for future predictions.

Better curated data makes AI Earth science models lean and quick. Google’s AI GenCast model can produce a 15-day forecast in eight minutes. A numerical weather model would take hours to produce the same forecast. That saved time equals saved energy and saved money. It would be tens of thousands of dollars cheaper to deploy GenCast over a numerical weather model.

Additionally, in the United States, half of the supercomputers that process conventional weather and climate models are housed at data centers in Florida that require significant cooling to keep them operational. This means that any way to lower the amount of work these supercomputers do creates magnified energy savings.

Terra incognita. Because the use of AI in research is still in its infancy, scientists and academic journals can sometimes be overzealous, publishing studies that might not be rigorous in their conception or methodology. AI models were largely developed by computer scientists; Earth scientists are still learning the best practices for their use. Lopez-Gomez notes that one way to preserve rigor in AI research in the Earth sciences is to have open and reproducible results that other researchers can check.

Another concern about the use of AI in climate research comes from how models are made. Because models are designed to recognize patterns or associations as opposed to working through scientific equations, they might get the right answers for the wrong reasons. Also, the use of AI in climate and weather modelling sometimes leads to “hallucination,” producing results that are impossible or nonsensical. But overall, the models are quite accurate and will likely become a fixture in the geosciences in the future. Their effect on climate and weather research will depend on how they are used.

“Let’s say that someone gives you a hammer when hammers were not a thing, and then you can decide what you do with that hammer,” Lopez-Gomez says. “You can build exciting stuff that helps others, or you can do harm.”