A modelling challenge to improve preparedness for highly pathogenicavian influenza epidemics: objectives, design and synthetic datageneration
This paper outlines the objectives, design, and synthetic data generation methods of the first-ever avian influenza modelling challenge, which utilized a simulated landscape of 9,160 poultry farms to engage 45 global teams in forecasting, analyzing transmission drivers, and evaluating control strategies for highly pathogenic avian influenza epidemics.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer
In the global effort to protect food supplies and public health, scientists rely on mathematical models to predict how diseases spread. These are not crystal balls, but rather sophisticated simulations that act like weather forecasts for epidemics. By feeding computers with data about how animals move, how viruses jump between them, and how humans intervene, researchers can test different strategies before a real crisis hits. This approach has become vital for diseases that move quickly, such as the flu strains that jump from birds to humans or livestock. The challenge lies in the fact that these models are only as good as the data they are tested against. In the real world, an outbreak is chaotic, messy, and often too dangerous to experiment on. To truly know if a model works, scientists need a perfect, known reality to compare it to—a scenario where they know exactly how the disease spread, where it went, and what would have happened if they had acted differently.
To solve this problem, a team of researchers created a completely artificial world called Jolly Island. This was not a real place, but a digital landscape designed to mimic the complexity of a major poultry-producing region. It contained nearly 9,200 farms, ranging from large indoor duck operations to smaller chicken broiler facilities, all scattered across a fictional archipelago. Into this digital ecosystem, the researchers introduced a highly pathogenic avian influenza virus. They let the disease run its course over several months, recording every single infection, every movement of birds, and every moment a farm was detected. This created a "ground truth," a perfect record of an epidemic that never actually happened. The researchers then invited forty-five independent teams of scientists from around the world to analyze this synthetic outbreak. The goal was to see if these teams could correctly predict the virus's path, figure out how it was spreading, and determine which control measures would work best, all while working with the same limited information that real-world officials would have during a crisis.
The results of this exercise, known as the HPAI Modelling Challenge, revealed both the power and the limitations of current scientific tools. The digital epidemic on Jolly Island behaved much like real outbreaks do. It started in a dense cluster of farms in the south, spread rapidly through local contact and the movement of birds, and eventually reached the northern parts of the island. The virus infected 560 farms in total, with the majority of cases occurring in the second stage of the outbreak. The researchers found that the spread was driven mostly by local transmission between neighboring farms, rather than by long-distance shipments or wild birds, although those factors did play a role. Crucially, the simulation showed that the speed of the response mattered immensely. In the early days of the outbreak, the system was overwhelmed; the capacity to cull infected birds was limited to just seven farms a day, causing a backlog that allowed the virus to spread further. As the simulation progressed, the capacity was increased to sixty farms a day, which helped bring the outbreak under control.
When the forty-five participating teams analyzed the data, they faced a difficult task. They were given information in stages, just as real officials receive data during an emergency. In the first phase, they only saw the beginning of the outbreak; in the second, the peak; and in the third, the resolution. The teams were asked to forecast where the disease would go next, identify the main drivers of transmission, and evaluate different control policies. The study found that while many teams could describe what had already happened, predicting the future remained difficult. The teams that used complex, mechanistic models—those that tried to simulate the actual biological and logistical processes of the virus—were generally better at inferring how the disease moved and what would happen if they changed their strategy. However, no single team had a perfect answer. The exercise highlighted that while individual models have value, combining the insights of many different teams often provides a more robust picture of the risk.
The paper concludes that these kinds of challenges are essential for preparing for future real-world outbreaks. By testing models against a known, synthetic reality, scientists can identify weaknesses in their methods without waiting for a real disaster to occur. The Jolly Island simulation proved that it is possible to create a realistic, high-stakes training ground for epidemiologists. It showed that while we can build sophisticated tools to understand disease, the human element of decision-making under pressure remains complex. The study does not claim to have solved the problem of avian influenza, but it has provided a new, rigorous way to train the people who will have to solve it. The data generated from this artificial island is now available to the global scientific community, serving as a benchmark to improve the tools that will one day protect our food systems and public health from the next great threat.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.