Verifiable abstention makes AI leak diagnosis accountable in water distribution networks
This paper introduces a verifiable abstention framework for AI-driven leak localization in water networks, where a physics-grounded executor and an LLM-audited supervisor collaborate to certify dispatches only when evidence meets strict contractual standards, thereby achieving high decision precision and accountability even under noisy field conditions.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Water utilities face a persistent and costly problem: treated water escapes through hidden cracks and broken pipes before it ever reaches a customer's tap. This "non-revenue water" represents a massive loss of resources and money, estimated in the billions annually. For decades, the solution has seemed to lie in better detection. Engineers have developed increasingly sophisticated computer models that use data from pressure sensors to guess where a leak might be. However, a critical gap remains between a computer's guess and a human's action. Digging up a street to find a leak is expensive, disruptive, and regulated; if a computer sends a crew to the wrong location based on a mere guess, the utility wastes thousands of dollars and damages public trust. The core challenge is not just finding the leak, but knowing with certainty when not to dig.
Researchers have now proposed a new way to handle this uncertainty, shifting the goal from simply predicting a location to making an accountable decision. Instead of forcing a computer to name a spot for every single anomaly it detects, the new system is designed to admit when it does not have enough proof to act. This approach treats the diagnosis like a medical differential diagnosis: rather than assuming every cough is pneumonia, the system considers other possibilities, such as a sensor malfunction or a sudden spike in water usage, and only recommends surgery if the evidence is overwhelming. By building a system that can confidently say "I do not know" and prove why, the researchers aim to bridge the gap between artificial intelligence and the real-world safety requirements of infrastructure management.
The team, led by researchers from several Chinese institutions, built a two-part digital system to test this idea on water networks. The first part, called the "executor," acts as a rigorous investigator. It takes a digital twin of a water network—a virtual replica that simulates how water flows and pressure changes—and tests competing theories against it. When a sensor detects a strange pressure drop, the executor does not just assume it is a leak. It simulates what would happen if that drop were caused by a leak, a sudden surge in demand, a broken sensor, or a valve that was closed by mistake. It runs these simulations to see which theory can actually reproduce the real-world data. If a theory cannot match the observed numbers, it is discarded. This process filters out false alarms before they ever reach a decision-maker.
The second part of the system, the "supervisor," acts as an independent auditor. It never guesses a location itself. Instead, it reviews the evidence package generated by the executor. This package contains only hard numbers derived from the physics simulations, such as how well a leak theory fits the data and how much better it is than the other theories. The supervisor checks these numbers against a strict set of rules, or a "contract," that must be met before any action is authorized. If the evidence is strong, the system issues a dispatch order with a digital certificate proving the decision was sound. If the evidence is weak or ambiguous, the system refuses to act, abstaining from sending a crew. This abstention is not a failure; it is a deliberate, verified outcome that prevents costly mistakes.
To test this approach, the researchers applied it to several different scenarios, ranging from computer-generated models to a real-world district network in China. In a standard computer simulation of a network with 381 junctions, a traditional system forced to guess on every event would be correct only about 32 percent of the time. The new system, however, chose to act on only about 40 percent of the events, but when it did act, it was correct 96 percent of the time. It successfully identified that many anomalies were not leaks at all, correctly distinguishing them from sensor errors or demand surges. When the system encountered a situation where the evidence was not clear enough to justify a dig, it simply abstained, avoiding the 68 percent error rate that plagued the forced-guessing approach.
The system was also tested on a third-party benchmark dataset that was generated by a different model, simulating a real-world scenario where the computer's internal map does not perfectly match reality. In this difficult test, the system abstained on 88 percent of the events, recognizing that the data was too noisy to be certain. For the four events where the evidence was strong enough, it acted with 100 percent precision, correctly identifying the leak location without a single false alarm. This demonstrated that the system could transfer its logic to new environments without needing to be retrained, a crucial requirement for real-world deployment.
The most rigorous test involved a real water district in China, using a register of 194 actual leak repair orders from 2025. In this real-world setting, the leaks were often very small, producing pressure signals so faint they were buried in the natural noise of the system. A traditional system forced to guess would have sent crews to dig 194 times, likely finding the correct spot only 12 percent of the time. The new system, adhering to its strict rules, decided that the pressure evidence was insufficient to justify excavation for almost all of these cases. It dispatched crews for only five events, and three of those were correct. While this seems like a low number, the researchers noted that the system was honest about the limits of the data.
To recover the remaining leaks that were too small for pressure sensors to detect, the system added a second layer of investigation. It looked at the total volume of water flowing into different districts of the network. Even if a leak is too small to cause a noticeable pressure drop at a sensor, it still adds to the total water entering a district. By comparing the water entering a district against the water leaving it, the system could identify which district contained the leak. Using this mass-balance approach, the system successfully identified the correct district for 85 of the 194 real leaks, achieving 100 percent precision at the district level. This two-tiered approach allowed the utility to target its resources effectively: sending a crew to dig only when the pressure data was clear, and sending a survey team to listen for leaks in the correct district when the pressure data was too weak.
The researchers emphasized that the system's strength lies in its ability to prove when it should not act. In a world where artificial intelligence is often expected to provide an answer for every question, this work suggests that the most valuable contribution might be the ability to say, "I cannot be sure." By grounding every decision in physics simulations and verifying them with independent checks, the system creates a defensible record for every action it takes. This accountability is essential for regulated infrastructure, where a wrong decision can cost millions and damage public trust. The study shows that by combining a physics-based investigator with a strict, rule-following auditor, it is possible to build an AI system that is not just smart, but also safe and reliable enough to be trusted with real-world decisions.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.