AI tool warns firefighters of oncoming explosive blaze

Firefighting is a risky job, particularly when fires in buildings can switch from bad to deadly in moments. The phenomenon of a flashover occurs when flammable materials in a room ignite almost simultaneously, producing a blaze which is only limited by the oxygen available. Identifying when this is about to happen – noting growing temperatures or flames rolling across the ceiling – is extremely difficult amid the many time-critical responsibilities firefighters have in the mayhem of a fire.

“I don’t think the fire services have many tools technology-wise that predict flashover at the scene,” said NIST researcher Christopher Brown, who is also a volunteer firefighter. “Our biggest tool is just observation and that can be very deceiving. Things look one way on the outside and when you get inside it could be quite different.”

The NIST researchers developed P-Flash (the Prediction Model for Flashover) to predict and warn of an incoming flashover; incredibly, it is designed to continue operating after heat detectors begin to fail, using remaining devices to monitor the blaze.

While models for predicting flashover based on temperature are not new, they have so far relied on uninterrupted streams of temperature data, which are obtainable under controlled lab conditions but are by no means guaranteed during a real fire. Heat detectors typically only work at temperatures up to 150°C - far below the 600°C at which a flashover begins. To bridge the gap created by lost data, NIST researchers applied machine learning.

“You lose the data, but you've got the trend up to where the heat detector fails and you've got other detectors. With machine learning, you could use that data as a jumping-off point to extrapolate whether flashover is going to occur or already occurred,” said NIST chemical engineer Thomas Cleary.

The researchers used temperature data from heat detectors in a sophisticated digital twin of a burning home to train their model as, for obvious reasons, it was not feasible to burn hundreds of homes to collect data. They ran more than 5,000 simulations with variations between each, such as the order in which pieces of furniture ignited and which windows and doors were closed.

The researchers trained the model on one batch of data and then tested it on the remaining batches, fine-tuning it based on its performance at predicting when a flashover was about to occur. They found that the model correctly predicted flashovers one minute beforehand for about 86 per cent of the simulated fires. Even when it missed the mark, it mostly did so by producing false positives, which is preferable to giving firefighters a false sense of security.

“You always want to be on the safe side. Even though we can accept a small number of false positives, our model development places a premium on minimising or, better yet, eliminating false negatives,” said NIST mechanical engineer and co-author Wai Cheong Tam.

After this, they trialled their model using real-world data produced during a recent study in a ranch-style home similar to their digital version. They found that their model was missing a phenomenon called the enclosure effect: when fires burn in small, closed-off spaces, heat has little ability to dissipate, so the temperature rises quickly. However, many of the experiments that form the basis of P-Flash’s training material were carried out in open lab spaces so temperatures from the real-world experiments shot up nearly twice as fast as the synthetic data.

Despite finding this weak spot in the tool, the researchers are encouraged and plan to represent the enclosure effect in their simulations through more full-scale experiments. They hope that eventually they will be able to provide a handheld tool for firefighters which communicates with detectors in burning buildings, providing firefighters with a vital safety tool.