The core idea borrows directly from nature. Real ants deposit pheromones along the trails they walk, and subsequent ants prefer paths carrying stronger chemical signals, a positive feedback loop that gradually steers the colony toward the shortest route to food. The ant colony optimization algorithm encodes this behavior in software: a population of virtual ants moves across a grid map, probabilistically choosing the next node based on pheromone concentration and a distance-based heuristic, then reinforcing the routes it traverses in proportion to their length. Over many iterations, the swarm converges on short paths. The method is prized for its robustness, distributed computing character, and adaptability, and it has already found use in robot navigation, transportation networks, and urban fire rescue.

But the standard algorithm has a blind spot when applied to burning villages. Its original heuristic measures nothing but geometric distance, which means it can happily route evacuees through an alley filled with lethal smoke simply because that alley happens to be short. Qiu and colleagues redesigned the heuristic to weigh three additional factors alongside raw distance: a human activity index that combines air temperature, toxic gas concentration, and crowd density through piecewise penalty functions; a passage difficulty coefficient that penalizes alleys narrower than three meters and effectively blocks those narrower than 1.5 meters with a penalty factor of 1000; and the remaining distance to the destination. The resulting equivalent evacuation length converts each grid cell into a risk-adjusted cost, so virtual ants automatically detour around high-temperature zones, CO-heavy corridors, and congested passages without any manual route instruction.

The fire physics feeding this risk map came from Pyrosim, a computational fluid dynamics tool. The team first built a three-dimensional model of Ma’an Village in Revit, capturing more than 30 ancient buildings, the road network, vegetation, and firefighting facilities in detail. They then imported that model into Pyrosim and simulated a five-megawatt fire ignited outside the village’s central drum tower over a 240 by 200 meter domain at half-meter resolution, some 6.1 million grid cells in total. Sensors placed at two meters in height at each exit recorded temperature, visibility, and smoke concentration. The simulation showed the fire source reaching 130 degrees Celsius within 50 seconds; by 500 seconds, the blaze had spread to nearby buildings, visibility had dropped below 13 meters across a hazardous core near the fire, and smoke concentration had climbed to 4.5 times 10 to the negative fifth mol per mol, well beyond the safety threshold.

The researchers extracted temperature, gas concentration, and crowd density data at two static snapshots, 200 and 500 seconds, and mapped them onto a 30 by 30 evacuation grid of 1.5 by 1.5 meter cells. Cell size matters here: the authors emphasize that grid cells represent the longitudinal step of a passage, not its transverse width, so genuinely impassable alleys are excluded by the passage difficulty coefficient rather than by grid resolution. This BIM-to-Pyrosim-to-MATLAB pipeline creates a closed data loop in which architectural geometry, fire field data, and route optimization flow automatically from one platform to the next, without repeated manual conversion, and the framework is explicitly designed as a generalized method for any historic dense settlement with similar characteristics, not just one village.

Three further upgrades sharpened the algorithm itself. An adaptive pheromone concentration steps down through the run, from 100 early on to 75 and then 50, boosting exploration at the start and refining convergence later. An adaptive evaporation coefficient gradually declines until it hits a floor tied to the fire’s impact level, with higher severity driving higher minimum evaporation to keep the swarm from fixating on stale routes. And a backtracking strategy rescues any ant that wanders into a dead end: the blocked node joins a taboo list and the ant retreats until it finds a passable cell, preventing infinite loops in the maze-like village geometry. After systematic single-factor tuning of the pheromone and heuristic weights, the optimal configuration settled at alpha equal to 2, beta equal to 9, evaporation factor 0.4, 70 ants, and 100 maximum iterations.

The performance gains were substantial and statistically solid. Running each parameter combination across 25 independent random seeds to eliminate stochastic noise, the team found that the improved algorithm shortened the shortest physical escape path from S1 to exit E4 from 17.75 meters under the standard algorithm to 15.11 meters under the improved version, a 14.9 percent reduction, the largest decrease among all seven tested exit paths. Every planned route came out shorter with the improved method, and paired-sample t-tests confirmed the differences were significant at p less than 0.05. Convergence improved even more dramatically: the upgraded algorithm locked onto optimal routes within an average of 3 to 10 iterations, while the standard version needed about 18, a gap significant at p less than 0.01, and its late-stage curve fluctuations indicated a stronger ability to avoid premature convergence on local optima.

When the simulated fire reached its spread phase, with 19 fire grids covering 42.75 square meters and four of seven exits blocked, the framework adapted cleanly, rerouting evacuees to the three remaining passable exits and still identifying exit E4 as the shortest at 15.11 meters. Computationally, the method is light: a single route calculation for the entire village takes roughly 28 seconds on an ordinary 16-gigabyte desktop with about 4.2 gigabytes of memory, and enlarged grids slow convergence slightly without stalling, suggesting the approach scales to larger heritage sites. The authors stress that all geometric parameters came from field measurements of the real village, grounding the virtual experiment in authentic conditions.

The team is candid about the work’s limits. The simulation covers only carbon monoxide as a toxic gas, excludes CO2, hypoxia, hydrogen cyanide, structural collapse, and wind effects, and relies on two static fire snapshots rather than continuous dynamic re-planning as occupants move. Validation so far rests on a single village and BIM-based co-simulation, with no field evacuation drills or historical fire records to calibrate the model, and the full route optimization currently targets physical distance alone without quantifying route safety. The authors also note that a complete, standardized fire risk ontology remains future work. Still, the framework offers heritage managers and emergency planners an immediate practical tool: an offline way to draft evacuation plans for the world’s many vulnerable historic settlements, with real-time sensor integration and dynamic occupant modeling as the clear next step toward protecting both lives and irreplaceable cultural heritage.

Subject of Research: Optimization of fire evacuation routes in historic dense settlements using BIM and an improved ant colony algorithm

Article Title: Generalized optimization method for fire evacuation routes in historic dense settlements based on improved ant colony algorithm and BIM

Article References: Qiu, J., Qin, J., Hu, Y., Liu, Y., & Qin, J. (2026). Generalized optimization method for fire evacuation routes in historic dense settlements based on improved ant colony algorithm and BIM. Discover Artificial Intelligence, 6(1), Article 1214. https://doi.org/10.1007/s44163-026-02256-2

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02256-2

Keywords: fire evacuation, historic dense settlements, ant colony optimization, BIM, path planning, Pyrosim, Revit, Ma'an Village, heuristic function, adaptive pheromone, emergency planning, cultural heritage