← News
Science Breakthroughs Science Breakthroughs Knowledge

The fastest evacuation route can leave more people unwarned

The fastest evacuation route can leave more people unwarned
More People Reached Safe Areas Model finding
More People Missed Warnings Trade-off
Fell About 60-Fold Warning spread
About 60-Fold warning spread drop

A route can get people to safety and still make an evacuation less informed. In simulations by Benjamin Jasperson and colleagues, adding conduits—guided paths through a modeled landscape—raised the share of agents reaching safe-zone minima, but also left more agents never exposed to the warning than unguided movement did. The route’s efficiency came with a social cost: people traveling along it encountered fewer neighbors who might otherwise have received the message (Jasperson et al., 2026).

That is a small but consequential shift in how to think about evacuation. A warning is not simply broadcast, then followed. It moves through people, and the people who hear it move through space. If those two processes alter each other, a plan optimized only for how quickly informed evacuees reach safety may overlook those who never learned there was danger.

The Science

The researchers built an agent-based model: instead of representing the population as a single average crowd, it tracks individuals, their locations, their information states and their movements. Each agent can be susceptible to information, exposed to it, informed, or reverted after the period in which they pass it on. The sequence is similar to an epidemic model, but the thing spreading is a warning rather than an infection.

An encounter occurs when the influence region around an informed agent overlaps the exposure region around a susceptible one. The susceptible agent becomes exposed immediately, then waits for a response delay before becoming informed. That delay represents the time people may need to interpret a warning, seek confirmation or decide to act. An informed agent eventually reverts and stops transmitting. The model allows these waiting periods and interaction radii to vary across agents, including through skewed Gamma distributions—probability distributions that can represent many shorter delays and a smaller tail of much longer ones.

Movement is driven by two forces. All agents feel short-range interactions with nearby agents, represented by a Lennard-Jones potential, a mathematical form that combines strong repulsion at very close distances with weaker attraction at somewhat greater distances. Only informed agents respond to the landscape-scale potential: modeled hazards are hills that push them away, while safe zones are valleys that draw them in. Conduits are sloping valleys that channel movement along a selected route. Random motion is added to the deterministic forces, so agents do not follow perfectly predictable paths.

(a) Gaussian-only potential landscape.
(a) Gaussian-only potential landscape. Source: Benjamin A. Jasperson, Logan Dehay

The key coupling is simple. A change in information state changes whether an agent follows the environmental landscape; a change in location alters which neighbors it can inform. The model’s central question is therefore not just whether people move toward safety, but how that movement changes the network of encounters through which the warning spreads.

The authors tested a sequence of scenarios, from a flat control landscape to landscapes with safe and hazardous regions, then with conduits. They also examined the gap between their spatial, encounter-based transmission and a well-mixed description. In a well-mixed model, every person is treated as if they had the same chance of meeting every other person. Real crowds do not behave that way: encounters depend on where people are, where they are going, and how long they remain near one another.

What They Found

First, adding potential minima changed who encountered the warning. Compared with an otherwise similar flat landscape, the landscapes with minima left far fewer agents never exposed. In the model, informed agents moved toward the minima and crossed paths with susceptible neighbors along the way. Their movement helped carry information into parts of the population that might not have encountered an informed person in a uniform environment.

This is the first important distinction in the results: “never exposed” is not the same as “exposed but still waiting.” The first group never had the modeled encounter that would begin the information process. The second had encountered information but had not yet completed its response delay. According to the study, the number left exposed was determined by the share of the population with very long delays; the number never exposed depended on how agents moved.

(d) t=0.55
(d) t=0.55 Source: Benjamin A. Jasperson, Logan Dehay

Then came the route trade-off. Conduits increased the fraction of agents reaching safe-zone minima compared with unguided evacuation. But more agents remained never exposed. The authors’ explanation is that a conduit shortens and constrains informed agents’ trips. They can reach safety more directly, but along the way they sweep past fewer susceptible neighbors. The same guidance that improves destination-finding can reduce opportunities to pass the warning on.

That result does not mean routes are harmful, or that planners should avoid them. It means that two goals—getting informed people to safety and extending the warning to everyone else—can pull in different directions. A route that scores well on one outcome may perform less well on another, especially when information spreads through local contact rather than reaching everyone at once.

The strongest numerical result concerns the limitations of the well-mixed assumption. The researchers estimated an effective transmission rate from the agents’ actual encounters. Within the first time unit, it fell roughly sixty-fold below the well-mixed value. In other words, a rate inferred under the assumption that everyone mixes freely can substantially overstate how quickly a warning spreads once the crowd’s spatial organization and state changes are taken seriously.

Spatial encounters rapidly suppress modeled information transmission

Relative transmission-rate scale, normalized to one-sixtieth of the well-mixed reference. The paper reports that its encounter-estimated effective rate falls roughly sixty-fold below the well-mixed value within the first time unit; these are normalized ratios, not measured rates with physical units.

Spatial encounters rapidly suppress modeled information transmission
LabelValue
Well-mixed reference60 relative scale
Spatial effective rate (within first time unit)1 relative scale

The paper attributes this decline to spatial mechanisms of net transmission suppression. As the process unfolds, state transitions and movement reshape who is near whom: informed people leave, some exposed people are still waiting to become informed, and others revert. A constant transmission rate—or even one adjusted only for the overall numbers in each state—does not capture that evolving geography.

(b) Radial drift vector field.
(b) Radial drift vector field. Source: Benjamin A. Jasperson, Logan Dehay

Why This Changes Things

Evacuation planning often separates communication from logistics. One set of tools asks whether the message reaches people; another asks how to move them through a road network or out of a hazard zone. This model suggests that the separation can hide feedback. Routes do not merely carry evacuees; they also determine which potential messengers encounter which people, and for how long.

That perspective matters most where official alerts are not the whole communication system. People may hear about danger from family, neighbors or colleagues, as the authors note in framing the problem. If some people rely on those nearby contacts, then who leaves, where they go, and how quickly they leave can affect who else learns of the threat. The model gives a formal way to ask whether a route, meeting point or staged departure pattern has consequences for those social links.

The striking sixty-fold gap also cautions against importing a familiar average-rate model without checking its assumptions. Mean-field models can be useful: they are comparatively simple and can reveal broad patterns in populations. But “average contact” can become a poor proxy when movement is goal-directed and information itself determines who moves. The authors’ result is not that mean-field models are always useless; it is that a constant or state-only transmission parameter may fail in the spatial regime they simulated.

There is a practical lesson here, though not yet a ready-made prescription. A route can be assessed on more than arrival rates. Planners could also ask how many people receive the warning before informed travelers disappear into safe zones, whether alternative channels reach isolated groups, and whether a route’s design leaves room for communication to continue. The answer might be better redundancy in alerts, strategically placed information hubs, or evacuation sequences that do not depend solely on neighbors passing the message. Those are possibilities to test, not interventions demonstrated by this paper.

The model also sharpens the meaning of “safe.” A person who reaches a designated safe zone is one kind of success. A person who remains in place without ever being exposed to the warning is a different kind of failure. Counting only those who arrive can obscure the people who never had the chance to respond. Evacuation metrics that track both outcomes could reveal trade-offs a single headline measure would miss.

What’s Next

This is a computational framework, not an evaluation of a real evacuation. The supplied paper text reports the qualitative comparisons and the approximate sixty-fold transmission-rate gap, but not the underlying counts, absolute transmission rates, population size, or detailed sensitivity results. The relative-rate chart above therefore normalizes the reported gap; it should not be mistaken for a measured rate or a forecast. The result depends on model choices—including local interaction radii, response-delay distributions, movement strengths and the shape of the potential landscape—and those choices will need comparison with observed human behavior.

People also do not move only because a mathematical landscape directs them. They interpret instructions, make decisions under uncertainty, follow leaders, use phones, and may change course when they see others doing so. The model captures some consequential features—local contact, delayed response, random motion, and information-dependent movement—but cannot establish how a particular crowd would behave in a particular emergency. Field data and carefully designed exercises would be needed to test whether the predicted route-versus-contact trade-off appears in practice.

A next step is to connect the model to realistic environments and communication systems: actual street layouts, hazard maps, alert reach, household networks and observed response delays. Researchers could then compare route designs not just by the number reaching safety, but by how many receive a warning, when they receive it, and who remains unreached. The authors also point toward coarse-grained descriptions, which might preserve the model’s spatial feedback without requiring every individual to be simulated.

The larger contribution is a question worth taking into emergency planning: when a warning makes people move, does the movement help the warning spread—or cut off its remaining paths? The answer will vary by setting, and this study does not settle it. But by making communication and motion part of the same model, it gives planners a way to look for the people a fast route might leave behind.