Wildfires have shaped the environment for millennia, but they are increasing in frequency, range and intensity in response to a hotter climate. The phenomenon is being incorporated into high-resolution simulations of the Earth’s climate by scientists at the Department of Energy’s Oak Ridge National Laboratory, with a mission to better understand and predict environmental change.

Two months into the 2023 peak summer fire season, Canadian wildfires had burned 25 million acres of land, disrupted the lives of millions and spread beyond the traditional confines of western Canada east to Nova Scotia. Smoke from the wildfires has drifted to heavily populated regions as far south as Georgia in the United States, across the Atlantic Ocean to Europe and into the Arctic Circle.

The impacts are being incorporated into large-scale simulations of the Earth’s climate, such as DOE’s Energy Exascale Earth System Model that reflects land processes like the carbon cycle for better predictions of the future climate. E3SM runs on the world’s fastest supercomputers, including the Frontier exascale system at ORNL, providing highly advanced simulations to better predict environmental change that could affect the energy sector.

ORNL scientist Jiafu Mao focuses on Earth system modeling, improving simulations of land surface responses and feedbacks to environmental change. The models evaluate synergies among historical fire data, carbon emissions, atmospheric factors such as temperature and precipitation, and major land variables such as vegetation condition, soil moisture and land use. His machine learning algorithms have supported better projections of wildfire and associated socioeconomic risk that can guide adaption and mitigation strategies.

Using AI to sharpen wildfire risk projections  

In one project, those algorithms were applied to improve the certainty of a series of Earth system models and predicted an increase in global wildfire exposure for the world’s population, gross domestic production and agriculture compared with untrained models. The research also indicated that models not constrained using the algorithms tended to overstate fire-related carbon emissions in regions with sparse vegetation. At the same time, the constrained models projected an increase in wildfire carbon emissions in tropical and subtropical regions as dense vegetation there dries and provides more fuel for fires. 

“We want to reach a better understanding and prediction of wildfire drivers, as well as potential vulnerabilities in terms of human health, ecosystem and infrastructure,” Mao said. The challenge is getting increased specificity in wildfire simulations from higher-resolution datasets. It would be helpful to gather data into a central repository that are now scattered among various federal, state, university and national laboratory sources needs to be gathered into a central location, he added.

“There are gaps in observational data, with much of the global wildfire record based on satellite products that started being collected and made available only about 20 years ago,” Mao said. “Long-term, high-spatiotemporal resolution, continuously gathered observations regarding the fires themselves as well as post-fire recovery processes are sparse.”

To enhance wildfire-related datasets, Mao and ORNL colleague Fernanda Santos have launched a Fire Community Database Network to encourage scientists and land managers to submit environmental data on burned areas to a central repository. Sharing such information can not only improve research, but also inform land management practices.

Forrest Hoffman, who heads the Computational Earth Sciences group at ORNL, is interested in the biogeochemical processes, including wildfire, that drive the evolution of the climate over multiple decades. He is laboratory research manager for the DOE Reducing Uncertainties in Biogeochemical Interactions through Synthesis and Computation Science Focus Area, or RUBISCO, which brings together scientists from national labs and universities to evaluate and improve Earth system models using laboratory, field and remote sensing data.

Wildfire has been traditionally underrepresented in Earth system models, an issue that Hoffman and his colleagues are working to address. “Getting the metrics right about burned areas derived from satellite remote sensing datasets means we can then better predict what will happen as climate change evolves under potential future scenarios,” he said.

Hoffman touted the machine learning methods Mao and other researchers are developing as part of the RUBISCO project as one way to get fire metrics right and represented.

Like Mao, Hoffman recognizes the need for new observational datasets. Sometimes researchers can’t see through the smoke on visible-sensor satellite images to evaluate fire emissions, and more multispectral and thermal imagery that provides finer detail on land surfaces would be helpful to fill in data on immediate and long-term impacts, Hoffman said.