Typically, flashovers occur when a structure fire causes the temperature in an enclosed space to reach at least 593 ºC (1,100 ºF) – this is the approximate autoignition temperature of many commonly used organic materials. As a result, they will emit flammable gases and ignite at more or less the same time, even if they haven’t been directly exposed to flames.

Needless to say, firefighters do not want to be in a building when a flashover occurs. Given the chaos and limited visibility inside burning structures, though, it can be difficult for them to spot visible warning signs such as flames rolling across the ceiling. And while many buildings are equipped with heat sensors, the devices usually stop working at a temperature of about 150 ºC (302 ºF), long before flashovers happen.

That’s where P-Flash – the Prediction Model for Flashover – is intended to come in.

Designed by scientists at the US-based National Institute of Standards and Technology (NIST), it utilizes machine learning algorithms that were trained on 4,033 varying computer simulations of a burning three-bedroom, one-story, ranch-style home. In most US states, this is the most common type of house.

Although the virtual heat sensors in that home cut out at 150 ºC, the algorithms learned to predict approximately when the temperature inside the home would reach 593 ºC, based on variables such as the rate at which the temperature was rising. When P-Flash was subsequently tested on additional burning-house simulations, it was 86 percent accurate at predicting flashovers one minute before they occurred. Additionally, when it was less accurate, it usually predicted them early … which is obviously better than predicting them late.