Details published in the Journal of Building Engineering reveal a new approach to fire evacuation that goes beyond simply identifying the nearest exit. Researchers at the National Institute of Standards and Technology (NIST) have developed Safe Step, an AI model that predicts fire outbreaks and dynamically updates safe evacuation routes within buildings. Unlike previous algorithms that focused on the shortest path, Safe Step takes into account changing conditions and directs occupants away from exits that may become unsafe if the fire spreads. “Fires can grow and spread,” said Hongqiang “Rory” Fang, a researcher at NIST and lead author of the journal paper. “Our model can help predict how a fire will progress and update emergency exit displays to direct people to the safest exit.” Some “smart” buildings are currently testing technology with dynamic exit displays that utilize real-time data from temperature and air quality sensors.
Safe Step: Fire spread prediction and evacuation using AI
Safe Step, a new artificial intelligence model developed by NIST, fundamentally changes traditional fire evacuation strategies by predicting the development of hazards in advance. This is a feature that building safety systems have never had before. Unlike algorithms that prioritize shortest paths, this model does not simply react to the current situation. Predict how a fire will spread and dynamically adjust recommended evacuation routes to maximize occupant safety. As described in the Journal of Building Engineering, Safe Step uses reinforcement learning. It leverages data from NIST’s established fire simulation tools and allows the AI to learn through simulated trial and error. The core of Safe Step’s effectiveness lies in its use of the fractional effective dose (FED) of a toxic gas, a metric that quantifies the severity of a hazard over time. The model consistently chooses the route that minimizes exposure to FED.
The researchers demonstrated the model’s advantages over traditional algorithms in test cases, highlighting in particular its ability to prevent occupants from unconsciously heading for exits, which could become dangerous as the fire spreads. In future iterations, we envision an AI system where each instance represents an individual occupant, allowing the model to account for crowding and adjust firefighter access. Wai Cheong Tam, a mechanical engineer at NIST, explained the team’s motivation: “We asked ourselves, ‘Can we build better algorithms that predict how fires will progress and save more lives?'” Pending regulatory approval and rigorous testing, widespread implementation is still five to 10 years away, but Safe Step represents a major advance toward intelligent fire suppression and proactive building safety.
Reinforcement learning guides safe evacuation route selection
Current evacuation algorithms often prioritize the shortest path, but this strategy has proven insufficient when building safety changes dynamically due to fire. Researchers are now leveraging artificial intelligence to predict fire development and guide occupants to truly safe exits. Unlike systems that rely solely on current sensor readings, Safe Step predicts how a fire will spread and allows evacuees to be rerouted in advance. When tested against traditional algorithms, Safe Step consistently identified safer evacuation routes, even for complex single-story structures. For example, as a fire spreads, the model can recognize that an initially closer exit may be at risk, and instead direct individuals to another exit that is further away, but ultimately safer.
Minimize hazardous exposure with fractional effective dose (FED) metrics
Researchers are increasingly focusing on quantifying exposure to hazards during fire evacuations beyond simple distance-based algorithms. This variable allows the “Safe Step” AI to prioritize routes that minimize cumulative risk, rather than focusing only on the shortest path to the exit. The lower the FED, the lower the risk to building occupants and provides a more nuanced safety assessment. Unlike previous approaches, Safe Step does not simply react to the current situation. Proactively predict how toxic gas concentrations will change as evacuees traverse the route. This predictive ability is essential because as a fire progresses, seemingly obvious exits can quickly become dangerous. The researchers demonstrated this in a test case involving a fire that spread through a hallway. While traditional algorithms guide users to the nearest exit, Safe Step can predict increasing danger and guide users to another exit that is further away, but ultimately safer. The implementation of FED in the model requires numerical values to determine the optimal route, and the effectiveness of the algorithm was verified through comparative tests with traditional methods using both simple and complex building structures.
“Although this work is still in the early stages of research and development, it represents an important step toward intelligent firefighting that can protect property and save lives through the effective use of advanced technology,” Huang said.
Future development of dynamic exit display and multi-agent system
The integration of artificial intelligence into building safety systems is moving beyond theoretical models to practical testing, with dynamic emergency exit displays being piloted in several “smart” buildings. These displays utilize real-time data from sensors that monitor temperature and air quality to assess exit safety and provide guidance to passengers, rather than simply showing the nearest route. This proactive approach represents a shift away from reactive emergency signs and is based on the “Safe Step” AI model developed by NIST researchers detailed in the Journal of Building Engineering. Researchers are already looking beyond single-layer applications, with ongoing research focused on adapting models to handle the complexity of multi-level structures. To more accurately simulate real-world scenarios, the team plans to create an AI system that uses multiple agents, each agent representing an individual building occupant.
The interaction between these agents allows the model to account for congestion, such as bottlenecks at building entrances, and dynamically adjust evacuation routes accordingly. This advanced coordination is not limited to occupant safety. This model could also facilitate easier access for firefighters. By directing evacuees to alternative exits, the system can clear the way for emergency responders and improve their ability to fight fires and assist vulnerable populations. NIST estimates that technologies like Safe Step could begin to be deployed in buildings within five to 10 years, but widespread adoption depends on regulatory approval and thorough reliability testing. “Although this work is still in the early research and development stage, it represents an important step towards intelligent firefighting that can protect property and save lives through the effective use of advanced technology,” concluded Huang, highlighting the potential of AI to fundamentally improve fire safety protocols.
“Fires can grow and spread,” said Hongqiang “Rory” Fang, a researcher at NIST and lead author of the journal paper. “Our model predicts how a fire will progress and helps us update emergency exit signs to direct people to the safest exit.”
