A deadly phenomenon known as "flashover" occurs when flammable materials in a room ignite almost simultaneously. A blind spot for firefighters, the event produces a blaze limited only by the amount of available oxygen.
A new tool — called P-Flash — estimates when flashover is imminent. The technology, built by researchers at the National Institute of Standard and Technology (NIST), also provides flashover warnings to responders.
What is a Flashover?
Flashovers are especially dangerous, because there are few warning signs to help firefighters detect them in advance. Some flashover indicators, like an increasingly intense heat or rolling flames across a ceiling, are easy to miss in the low-visibility, high-stress environment of a rescue.
“I don't think the fire service has many tools technology-wise that predict flashover at the scene,” said NIST researcher Christopher Brown , who also serves as 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 Prediction Model for Flashover, or P-Flash, pulls data from an array of nearby heat detectors, including those in adjacent rooms, to recover temperature data from the room of fire origin and estimate potential for flashover.
What is P-Flash?
The NIST-developed model predicted imminent flashovers in over a thousand simulated fires and more than a dozen real-world fires. Experimental evaluation, just published in the Proceedings of the AAAI Conference on Artificial Intelligence , suggests the model shows reliable prediction in anticipating simulated flashovers.
According to the report, the model performance is approximately 83% and 81%, respectively, for current and future flashover occurrence, considering heat detector failure at 150 ̊C.
Heat detectors, which are commonly installed in commercial buildings and can be used in homes alongside smoke alarms, are for the most part expected to operate only at temperatures up to 150 ̊C (302 degrees Fahrenheit), far below the 600 ̊C (1,100 degrees Fahrenheit) at which a flashover typically begins to occur. To bridge the gap created by the missing data, NIST researchers applied a form of artificial intelligence known as 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, a co-author of the study.
Burning Down the House (...Virtually)
Machine-learning algorithms use large quantities of data to predict outcomes. To get a large amount of information about house fires, however, requires a digital dwelling: a simulation of a burning three-bedroom, one-story ranch-style home.
To build P-Flash, Cleary and colleagues fed their algorithm temperature data from heat detectors from the virtual house — the most common type of home in a majority of states. The team burned this virtual building repeatedly — running 5,041 simulations, in fact— using NIST’s Consolidated Model of Fire and Smoke Transport, or CFAST , a fire modeling program validated by real fire experiments.
Each of the 5,000-plus simulations had slight, but critical variations. Windows and bedroom doors were randomly configured to be open or closed. Furniture came and went, and moved around. The front door opened and closed.
Heat detectors placed in the rooms produced temperature data until they were inevitably disabled by the intense heat.
To learn about P-Flash’s ability to predict flashovers after heat detectors fail, the researchers split up the simulated temperature recordings, allowing the algorithm to learn from a set of 4,033 while keeping the others out of sight. Then, the team quizzed P-Flash on 504 simulations, tweaking the model based on its guesses.
The researchers found that the model correctly predicted flashovers one minute beforehand for about 86% of the simulated fires. Many of the misses, according to the team, were false positives, which predicted the flash at an inaccurately early moment but at least did not provide firefighters with a false sense of security.
Testing with Real Data (and Real Fires)
Additionally, NIST tested P-Flash further by comparing its predicted temperature data to temperatures measured in 13 real house fires, purposefully lit during Underwriters Laboratories (UL) experiments.
With the temperature data from the UL experiments, P-Flash, attempting to predict flashovers up to 30 seconds beforehand, performed well when fires started in open areas such the kitchen or living room. When fires started in a bedroom, behind closed doors, however, the model could almost never tell when flashover was imminent.
The team identified a phenomenon called the enclosure effect as a possible explanation for the sharp drop-off in accuracy. When fires burn in small, closed-off spaces, heat has little ability to dissipate, so temperature rises quickly — more quickly than the fires in the open lab spaces that provided P-Flash's early training data.
The researchers’ next task is to perform more full-scale experiments that zero in on the enclosure effect and represent it in simulations. With improvements, the team hopes to embed the system in handheld devices that communicate with detectors in a building through the cloud, notifying responders of danger spots and when it's time to get out.
In an email interview with Tech Briefs, NIST engineer Thomas Cleary, explains more about when he expects firefighters to be able to use the model. Cleary answered in collaboration with his colleagues Christopher Brown, Jonathan Griffin, Andy Tam, and Anthony Putorti.
Tech Briefs: How do you “burn a virtual building?” That seems like a very interesting task. What are you changing on the building each time? And how does that inform your model?
Thomas Cleary: A model like P-Flash is trained using large datasets from a range of fire scenarios. It’s unrealistic to generate the necessary amount of data from real fires so we use computer fire models. Specifically, the NIST fire model, CFAST, is used to simulate fires in a modeled “virtual” building.
For a fixed building layout, we include a wide range of fires, from slow to ultra-fast growing fires, and vary their locations, and the vent opening conditions (i.e., doors and windows) to mimic what is plausible in real fires.
Approximately 5000 simulated fires with flashover occurrence are used to train P-Flash such that it learns the useful trends and patterns to correlate flashover conditions to the limited temperature information.
Tech Briefs: What inspired this idea? What is the current technology available to help a firefighter address flashover?
Thomas Cleary: The inspiration for our current research stems from previous research  investigating sending the state of the fire alarm control panel and information from smoke and heat detectors, to the fire service while on route to a fire so they have a sense of fire location and spread prior to arrival. A natural extension is to use the data from detectors in a predictive fashion to provide forecasting. Other research at NIST using the fire model CFAST in Monte Carlo modeling of fire scenarios suggested the large data sets for machine learning/AI are readily attainable from computer fire modeling.
Currently, firefighters rely on their senses, training, or at best handheld thermal sensors or thermal imaging cameras to get an idea of possible transition to flashover. Unfortunately, one would need to be at or near a room that is approaching flashover to have a chance to recognize the danger.
Tech Briefs: What have you heard from firefighters about their challenges with flashover?
Thomas Cleary: Currently, firefighters try to avoid flashover based on their experience with interpreting observational clues of flashover, such as rollover, high heat, etc., within the building structure and dark smoke coming out from the exterior windows. However, the transition to flashover is typically within seconds and, in general, the flashover indicators are not easy to recognize and if missed, would put lives in danger. We hope that our work will enhance experience-based firefighting by facilitating data-driven firefighting.
Tech Briefs: How do you turn the model into a usable tool? Can firefighters use this model right now?
Thomas Cleary: The focus of the research was to rely on building data that is or could easily be provided from available building sensors. One way to translate the research into reality is to integrate the model into a smart fire alarm control panel that would gather the temperature data from installed heat detectors and includes a computer module that can process the data and make the real-time predictions. From the fire alarm control panel or other suitable piece of equipment, the prediction would be sent to the incident commander, or individual firefighters if deemed suitable. The exact mechanism of providing such predictive analytics is not decided and would require input from the fire service to develop a consensus.
Firefighters can’t use the model now. Before the model can be developed and incorporated into a smart fire alarm control panel, we feel we need to verify model performance (real-time prediction) in building fire tests with heat detectors.
Tech Briefs: What’s next regarding this work?
Thomas Cleary: We are currently extending P-Flash to work for different building layouts. In the next year or so, we are planning demonstrations for building fire experiments, and have begun to engage with fire safety (alarm) equipment manufacturers about the model’s capabilities.
What do you think? Share your questions and comments below.
 Reneke, P. A. (2013). Towards Smart Fire Panels . NIST TN 1780. US Department of Commerce, National Institute of Standards and Technology, MD.