Construction and Application of an AI-Agent-Based Workflow for the Rapid Survey of Building Damage and Functionality

When the ground shakes, the immediate question for a city is not just whether a building stands, but whether it can still be used. A hospital might remain structurally sound, with its walls and beams intact, yet be completely useless if its medical equipment is shattered or its plumbing systems severed. For decades, engineers have assessed buildings by looking at individual parts: a cracked wall, a broken window, or a bent beam. They assign a grade to each piece of damage, but this approach often fails to answer the most critical question for survivors: is the room inside safe and functional? The gap between knowing a wall is cracked and knowing a hospital operating room is unusable has long been a difficult problem to solve, relying heavily on the subjective judgment of human inspectors who must piece together a complex picture from scattered clues.

Researchers Xu Zheng and Jin Liu have proposed a new way to bridge this gap, creating a digital workflow that acts like a team of specialized assistants to evaluate a building after an earthquake. Instead of a single computer program trying to do everything at once, they built a system where different artificial intelligence agents work together, each with a specific job. One agent looks at photos taken on the ground to spot damage. Another agent consults a digital library of building codes to understand what that damage means. A third agent uses a probabilistic model to figure out how damage to a single wall might affect the ability of an entire room to function. Finally, a fourth agent compiles all this information into a clear, standardized report. This system is designed to move beyond simple damage counting, translating physical cracks and broken pipes into a precise assessment of whether a building can still do its job.

The core of this new method is a shift in perspective. Traditional methods often treat a building as a collection of parts, where the worst damage simply dictates the final verdict. If a single critical component is broken, the whole system is often declared failed, regardless of whether other parts are working. The researchers argue that this is too blunt an instrument. In their system, the AI agents work to distinguish between physical damage and functional loss. They use a network that maps how different parts of a building depend on one another. For example, in a hospital operating room, the air filtration system, the gas pipelines, and the medical machines are all vital. If the walls are slightly cracked but the gas lines are intact, the room might still be usable, perhaps with some limitations. However, if the gas lines are severed, the room becomes useless even if the walls are perfect. The system calculates these relationships to provide a nuanced answer rather than a simple pass or fail.

To test this idea, the team applied their workflow to a realistic scenario involving a hospital operating room. They fed the system photos and data describing various types of damage, such as cracks in the walls or displacement of the ceiling. The system then traced these observations through its network of agents. The perception agent identified the damage, the knowledge agent matched it to specific building code rules, and the inference agent calculated the impact on the room's overall function. In one specific case, the system found that while the non-structural parts of the room, like the walls and ceiling, were moderately damaged, the critical medical equipment and gas systems were intact. The workflow concluded that the room was degraded but still operable. This is a crucial distinction. A simpler method that just looked at the worst damage would have likely declared the room completely unusable, potentially leading to unnecessary evacuations or the abandonment of a facility that could still save lives.

The researchers also explored what happens when the damage is more severe. They ran a series of scenarios where they changed the type and location of the damage to see how the system responded. They found that the system correctly identified that severe damage to non-structural elements, like a collapsed ceiling, could make a room unsafe for anyone to enter, even if the medical equipment was fine. Conversely, they found that severe damage to a single functional system, like the medical gas supply, would render the room unusable for its intended purpose, even if the building structure was perfect. The system successfully separated these two types of failure, showing that a building's ability to function depends on the specific combination of damaged parts, not just the total amount of damage. This ability to tell the difference between a building that is physically unsafe and one that is functionally broken is the primary advance of their work.

A key feature of this workflow is its ability to know when it is unsure. The system is designed to flag cases where the evidence is unclear or the damage is complex. When the AI agents encounter a situation where their confidence is low, or where different parts of the system give conflicting results, they automatically pause and ask a human expert for help. This creates a partnership between the machine and the human. The machine handles the routine work of scanning thousands of photos and checking them against codes, while the human steps in only for the difficult cases that require judgment. In their tests, this approach allowed the system to process most scenarios automatically while ensuring that the most critical decisions were reviewed by a person. This balance helps to keep the error rate low without slowing down the entire assessment process.

The team also compared their new method against older, more traditional ways of assessing damage. They found that the old methods, which simply added up the worst damage found in any part of the room, tended to overestimate the loss. They would often declare a room completely lost when it was actually only partially damaged. The new workflow, by contrast, provided a more accurate picture of the building's actual capabilities. It showed that a room with a cracked wall but working medical systems is different from a room with a working wall but broken gas lines. This level of detail is something that standard checklists often miss. The researchers demonstrated that their system could produce a detailed, auditable report that explained exactly why it reached a certain conclusion, citing the specific building codes and evidence that led to the decision.

Despite these promising results, the researchers are careful to note the limits of their current work. The system was tested using a mix of real photos and expert-provided data, but it has not yet been deployed in a real-world disaster zone where conditions are chaotic and unpredictable. The part of the system that identifies damage from photos, while effective, still makes mistakes, particularly when estimating the size of cracks or when the photos are taken from difficult angles. The researchers acknowledge that the system is currently a tool for rapid screening and triage, meant to help engineers prioritize which buildings need immediate attention, rather than a replacement for a full, formal engineering inspection. They emphasize that the final responsibility for safety decisions must always remain with licensed human professionals.

The path forward for this technology involves refining the AI's ability to see and measure damage more accurately and expanding the system to handle different types of buildings and different building codes from around the world. The researchers plan to test the system with more diverse sets of images and to improve the way it handles uncertainty. They also aim to integrate the system more deeply with the process of updating building codes and training new engineers. The ultimate goal is to create a tool that can be used by first responders and engineers in the chaotic hours after an earthquake to quickly determine which buildings are safe to enter and which are not. By turning a complex, subjective process into a structured, evidence-based workflow, this research offers a new way to think about resilience, moving the focus from simply counting broken parts to understanding what a building can still do for the people who need it.