Assad Oberai had recently moved to Southern California from upstate New York when the Bobcat Fire broke out in September 2020. As he watched, smoke choked the region and flames lit up hillsides closer and closer to his home in La Cañada Flintridge, just north of Los Angeles. The fire seemed to have its own internal logic. Oberai wanted to figure out what it would gobble up next—an inviting puzzle for a mechanical engineer.

He did some research and discovered that traditional fire path predictions stem from photos taken by polar-orbiting satellites and infrared technology on planes that gather data on the fire’s perimeter. Both fly over a particular spot on the planet occasionally throughout the day and send information back to officials, creating a sporadic snapshot of the disaster, made more complicated by smoke that can obscure parts of the pictures. Oberai wanted to find a better way to divine a fire’s future.

Navigating the scale and complexity of global problems—predicting the outcomes of climate change, anticipating biodiversity loss—seems to boggle the human mind. Meanwhile, artificial intelligence and machine learning are increasingly able to see patterns as they happen, and to use them to predict the future. A deep understanding of the natural world in flux may require that humans use millions of machines to decode the environment.

A deep understanding of the natural world in flux may require that humans use millions of machines to decode the environment.

Oberai’s previous research at the University of Southern California re-created the messy physical world with elaborate computer models: diseases that spread through the body, turbulent air that flowed over airplane wings, and cells that multiplied within tumors. He wanted to understand that same messy world in the context of fire’s behavior. He and his colleagues created a computational model that uses additional satellite imagery and more sources of data than the traditional predictive models do, including information about topography, weather, and wind currents. The group trained the algorithm with data from four real wildfires that took place between 2020 and 2022 and found that it was successful at predicting their behavior.

The model augurs several outcomes and the likelihood of each one in a way that is similar to weather or hurricane prediction. By understanding exactly what the current state of a fire is, it’s easier to predict where it will go next. NASA now funds the project.

Computer science and environmental research are merging in ambitious ways. Scientists have historically tracked tree stress and mortality through ground surveys, limited to what humans can physically access. In January 2024, a group of researchers published a study that used machine learning and aerial photographs to count dead trees in California, cataloging 91.4 million—potentially a massive source of fuel for future forest fires. Similar research collaborations focus on reducing emissions, catching poachers, and even developing an “internet of animals.”

It makes sense to turn to machines to monitor life on Earth. Climate change is happening at a speed beyond what human-paced research can manage, and earth science is all about data—data coming from satellites, from sensors in streams, from camera traps. There simply aren’t enough humans to track and analyze it all. But machine learning and AI algorithms are “data hungry,” Oberai pointed out. AI moves fast enough to potentially be an invaluable ally to us.

On the other coast, Priya Donti, an assistant professor at MIT, also brings together researchers in the climate science and AI worlds through the nonprofit research collective Climate Change AI, which she cofounded in 2019. She became inspired to work on the climate crisis in a biology class as a high school freshman, but her computer science classes also tugged at her. She set out to bridge the two.

Donti sees the areas where AI can aid humans right now, such as optimizing freight-trucking routes to reduce emissions. Her own research examines how AI can improve power grids, which must balance fluctuating renewable power coming in with power going out to charge batteries. Outsourcing that management to AI can make human-designed systems work smarter.

From her office in Cambridge, Massachusetts, Donti described a project funded by Climate Change AI that identifies aquaculture sites with high potential for more sustainable production in Indonesia and the Philippines. Shrimp farming has driven 30 percent of the loss of mangroves in the area. Combining satellite data with machine learning, researchers created an interactive map to identify places that shrimp aquaculture could be located—to both conserve mangroves and produce shrimp to support local livelihoods and aid food security. The tool they created is now also being deployed in Ecuador.

AI, while powerful, is not a silver bullet. The same methods used to estimate methane emissions from cows or identify butterfly biodiversity can be used by fossil fuel companies to increase efficiency in drilling or to generate science misinformation.

By 2030, AI is also poised to send data-center power usage skyrocketing, well over double today’s consumption—and it’s mostly fossil fuels generating the power to meet that rising demand. The best path forward, Donti said, is for computer scientists to pick AI algorithms that are the right size for the problem they are trying to solve, as larger models eat up more energy. It may also require that average users make smarter choices about when to use AI. “If AI is used to decarbonize the power system, that presumably is a relatively different use case than AI being used to generate an image for a slide,” she said.

Back on the West Coast, Oberai knows there will be more fires. He hopes the models he’s pioneering can help people make better decisions on the ground when those emergencies are evolving quickly. “One can come up with really good, accurate models driven by machine learning and artificial intelligence in this particular area and make an impact,” he said. “It’s a natural marriage.”