USC and Cal Fire Build GenFire, an AI Tool to Trace Where Wildfires Start

Why finding a fire's origin is slow

The project follows a long investigation into the Eaton Fire. After almost 20 months, the Los Angeles Fire Department recently released a report identifying the source of that fire. According to the USC Viterbi School of Engineering, such timelines are not unusual, and identifying where a fire started remains a major challenge for fire experts and scientists.

Traditional investigations rely on physical evidence such as scorch marks. That requires investigators to be on site and can take a long time. Limited satellite data and a lack of tools grounded in scientific rigor make real-time tracking harder still.

The team says terrain, weather and fuel types constantly shape a fire's direction and speed, which makes it hard to tell where a fire began from the areas it reached later. Most satellite systems observe at intervals of 12 to 24 hours, so a fire may be "raging for six, 10, 12 hours" before a satellite captures it. ALERTCalifornia, a network of cameras on hilltops and mountaintops, offers high-resolution video in near real time, but the cameras are limited by distance and terrain obstructions.

Peter Beerel, one of the researchers, pointed out that existing approaches have not relied on a consistent scientific framework. Previous methods have reached an average accuracy of about 2 kilometers from the initial fire spot, which leaves investigators too much room.

How GenFire is meant to work

GenFire is a generative AI tool. The team plans to train it on environmental data, including wind, temperature, humidity, terrain and vegetation, along with observations of where and when a fire was detected. The tool will work backward from a fire's observed spread and generate possible scenarios for how it developed during the first 24 hours. By comparing those scenarios, GenFire will identify the most probable ignition area and reconstruct the sequence of early growth.

The researchers also plan to train the model on thousands of virtual wildfires produced by a physics-based fire simulation model, so the AI can learn how environmental factors influence fire behavior. Afterward, the team plans to validate it against approximately 100 historical wildfires with known origins, comparing predicted ignition locations with the real ones to measure accuracy.

The team also intends to build on FireLoc, a wildfire-specific geolocation framework Beerel previously introduced in his work on wildfire crowdsensing. It would be adapted to process video from the ALERTCalifornia and ALERTWest networks, turning continuous footage into time-stamped maps of fire activity and giving GenFire observations at intervals of minutes.

Project lead Assad Oberai described the approach this way: "We are using forensic data reconstruction to look backward and determine precisely where that spread began."

Team, testing and open source plans

Oberai chairs USC Viterbi's Department of Aerospace and Mechanical Engineering and is a Hughes professor of aerospace and mechanical engineering. The team also includes Beerel, a USC electrical and computer engineering professor, and Shawn Zimmermaker, Cal Fire Deputy Chief and chair of the Wildland Fire Investigation Subcommittee (WFISC).

The project spans two years. The team will collaborate with WFISC to test GenFire and ensure the tool's operational relevance. The researchers also plan to make GenFire open source, so fire investigators and scientists can use, evaluate and build upon it.

What it could mean for investigators

The researchers aim to speed up and standardize fire investigations and to identify a fire's source more precisely, supporting prevention efforts. Oberai said there is "a significant legal and scientific need for this work," adding that identifying a fire's origin provides the transparency and accountability needed to distinguish between natural causes, accidents and criminal acts such as arson.

According to the team, GenFire could help investigators better understand ignition events, including potential human or infrastructure-related causes, and inform future wildfire prevention and mitigation. The tool is still in development, and its accuracy has yet to be measured against historical fires.