Firefighting has evolved tremendously since bucket brigades, where people passed buckets of water hand-to-hand from a well or river to the fire.
Across the U.S., researchers are building AI systems that can interpret fire behavior, chart safer evacuation routes, analyze structural vulnerabilities, and monitor firefighters’ health in real time. The technology is not replacing human judgment; it’s augmenting it—offering insights that are otherwise impossible to capture in the middle of chaos.
Among the multitude of tests, measurements, and investigations carried out by the U.S. Department of Commerce’s National Institute of Standards and Technology (NIST), it studies how fires start and spread under various scenarios. The agency’s Artificial Intelligence Enabled Smart Firefighting program, led by researchers Wai Cheong “Andy” Tam and Anthony Putorti, brings together machine learning, fire dynamics, and physiological research.
NIST’s approach begins with gathering high-quality data from test burns, sensors, wearable devices, building layouts, and firefighter training scenarios. That information is refined and fed into models that are trained, validated, and then deployed in controlled environments for testing. The goal is to produce predictive tools, and tools that can shift in real-time to the unpredictable conditions of actual fires.
One effort focuses on a notoriously dangerous moment. When a room reaches critical heat, everything ignites nearly simultaneously, called a flashover. Firefighters rely on experience to sense warning signs, but in a large or compartmentalized structure, the indicators can be subtle. NIST is developing machine learning tools that can forecast flashovers inside such residences or buildings. The system reads heat signatures, smoke patterns, and environmental data to alert crews before a flashover.
Escape routes are also getting scrutinized through AI. Traditional evacuation plans rely almost entirely on identifying the shortest route to an exit, assuming the fire behaves predictably. It rarely does. When flames move faster than expected or smoke fills a hallway, evacuees can find themselves trapped. Using reinforcement learning, NIST researchers are building dynamic evacuation models that update as conditions deteriorate, steering people toward the safest available path. Instead of static instructions posted on a wall, the AI behaves like a guide that adapts on the fly.
Lithium-ion batteries present a different kind of threat—fast, explosive, and often silent until it’s too late. Mr. Tam and Mr. Putorti developed a method to detect the distinctive click-hiss of a battery’s safety valve as it breaks under pressure. That instant of expanding gas is the precursor to a violent thermal reaction.
“While watching videos of exploding batteries, I noticed something interesting,” Mr. Tam said. “Right before the fire started, the safety valve in the battery broke and it made this little noise. I thought we might be able to use that.”
They taught the software by using more than 1,000 unique audio samples of what a breaking safety valve sounds like. The algorithm works remarkably well. Using a microphone mounted on a camera, the researchers detected the sound of an overheating battery 94% of the time.
“I tried to confuse the algorithm using all kinds of different noises, from recordings of people walking, to closing doors, to opening Coke cans,” Mr. Tam said. “Only a few of them confused the detector.”
The ability of AI to detect that sound can give residents, firefighters, and building systems an early warning. With lithium-ion batteries powering everything from scooters to cars to home electronics, that capability to prevent, reduce, or respond to fires quicker, is massive.
Another way AI is being applied is to one of the fire service’s most persistent killers: sudden cardiac death. Over the years, it has accounted for over 40% of on-duty firefighter fatalities. Many of those who died had passed their medical evaluations with no signs of risk. The problem isn’t lack of fitness; it’s the stress of the work. Extreme heat, heavy gear, unpredictable bursts of exertion, and poor visibility can strain even healthy hearts.
NIST researchers and medical partners developed a heart health monitoring system called H2M, using 24-hour electrocardiogram (ECG) data from 112 career firefighters. The model, built using convolutional neural networks, can classify normal, abnormal, and noisy heart rhythms in real time using six-second ECG segments. It outperformed several state-of-the-art models and demonstrated that training on real firefighter data is essential—using ECGs from non-firefighters reduced accuracy by about 40%. The goal is a wearable system that warns firefighters when their physiology begins to deteriorate, giving them time to step back before a fatal event occurs.
At the AI in Fire Engineering Summit this spring at the University of California, Berkeley, teams presented machine learning models that could replace lengthy structural fire simulations with faster, highly accurate predictions. Others demonstrated wildfire damage prediction tools trained on California’s inspection datasets, revealing which structures face the highest risk and why. These models can strengthen wildfire preparedness, and invigorate mitigation efforts.
Fire science has traditionally been rooted in post-incident analysis—studying how fires behaved after they were extinguished, but with AI, there is a change in that philosophy.
