water · infrastructure · family: the model was never trained on this
two million miles of pipe, no failure map
Cities cannot predict which of their millions of buried water pipes will fail next
Problem statement
The United States has approximately 2 million miles of buried water pipes, most installed after World War II and now approaching or exceeding their designed lifespan. The EPA estimates that bringing these systems to a state of good repair will cost $625 billion over the next 20 years, but cities cannot spend that money rationally because they don't know which pipes are most likely to fail. ASCE reports approximately 240,000 water main breaks per year, resulting in roughly $2.6 billion in repair and maintenance costs. The fundamental problem is that pipes are buried, their condition is invisible, and the historical data that would enable failure prediction — installation date, material, soil conditions, pressure history, repair records — is incomplete, inconsistent, or locked in paper records across decades of municipal administration.
Why this matters
Water main breaks cause immediate public safety hazards (sinkholes, flooding, contamination), disrupt transportation and commerce, and damage surrounding infrastructure. The median age of the U.S. water workforce is 48, and the EPA found in 2020 that approximately one-third of the drinking water and wastewater workforce will become eligible to retire over the decade, taking institutional knowledge of buried infrastructure with them. McKinsey estimates that waste and water infrastructure globally requires $6 trillion in investment through 2040. The problem is not confined to the U.S. — the MDPI study puts average water losses in Europe at roughly 26%, and lists Belgium, Austria, the Netherlands, Denmark, Germany and Luxembourg, and separately Australia, as losing less water than the U.S. Replacement remains partly reactive: ASCE says predictive tools are reducing reactive replacements, but reports that only about 30% of utilities have fully implemented an asset management plan, and the MDPI study calls for proactive rather than reactive approaches to managing distribution systems.
What’s been tried and why it hasn’t worked
ASCE reports that many utilities are adopting asset failure prediction technologies, which improve the ability to identify issues before they become failures, but that only about 30% of utilities have fully implemented an asset management plan. Pipe age and material alone do not settle failure risk; site conditions matter too: ASCE reports that ductile iron pipes in highly corrosive soil break at six times the rate of those in less corrosive soil. Acoustic leak detection — listening for the acoustic signature of a leak with noise loggers or handheld listening sticks — is the conventional method for finding active leaks but is reactive (finds leaks after they start) rather than predictive, and the handheld listening sticks are used by highly trained technicians. AI-based satellite leak detection, a newer approach, uses satellite radar imagery and machine learning to identify subsurface moisture anomalies, narrowing the pipe length that must be physically inspected; a 2025 study using Atlanta water-audit data found both approaches economically favorable over a 3-year period, with conventional acoustic detection generating USD 2.4 million in net benefits, 50% more than the satellite approach.
What would unlock progress
A low-cost, scalable approach to estimating pipe condition without direct inspection would transform water infrastructure management. This could combine multiple indirect data sources — soil type from geological surveys, traffic data from transportation departments, historical repair records (even incomplete ones), water pressure data from existing SCADA systems, weather patterns, and satellite/aerial imagery — into a machine learning model that predicts failure probability at the pipe segment level. The key insight is that even a rough risk ranking that correctly identifies the top 10% highest-risk pipes would dramatically improve capital allocation compared to the current reactive approach. Utilities need tools that work with the incomplete, messy data they actually have, not tools that require comprehensive digital asset inventories they'll never build.
Entry points for student teams
A student team could build a pipe failure risk model for a specific city using publicly available data: water main break reports (many cities publish these as open data), soil maps from USDA Web Soil Survey, street and traffic data from OpenStreetMap, and pipe age/material data from utility GIS layers (some cities publish these). The team would train a spatial model to predict break probability and validate it against held-out break data. Even a model that modestly outperforms age-based prediction would demonstrate the value of multi-source data integration. Skills in geospatial analysis, machine learning, and civil/environmental engineering would be most relevant. A more design-oriented team could prototype a mobile app for utility field crews to capture pipe condition data during routine work, building the asset database that currently doesn't exist.
Genes
Structural cousins: stuck the same way, in other fields
Sources
Alastair Green, Ishaan Nangia, and Nicola Sandri, "The infrastructure moment," McKinsey, Report, September 9, 2025. (accessed 2026-02-12). Supplemented with "The Global Risks Report 2025," WEF, Insight Report, January 2025; ASCE 2025 Infrastructure Report Card, Drinking Water category, (accessed 2026-09-30); and Prashant Nagapurkar, Naushita Sharma, Susana Garcia, and Sachin Nimbalkar, "Evaluating Acoustic vs. AI-Based Satellite Leak Detection in Aging US Water Infrastructure: A Cost and Energy Savings Analysis," *Smart Cities*, MDPI, 8(4):122, 2025. go to source 1 ↗ go to source 2 ↗
Working notes
Notes made while writing and checking this brief, shown as written. They use working shorthand.
- McKinsey's "The Infrastructure Moment" (September 2025) provides the investment context. The $625B EPA estimate is from the most recent Drinking Water Infrastructure Needs Survey and Assessment.
- The 2025 MDPI study comparing acoustic vs. satellite leak detection in Atlanta is a valuable reference for understanding the current state of detection technology and its limitations.
- Cross-domain connection: this problem shares structure with `infrastructure-cascading-failure-modeling` — water main breaks can cascade to road failures, building damage, and service disruptions. Also related to `energy-grid-battery-scale-failure-prediction` in that both involve predicting failure in large, distributed infrastructure networks with heterogeneous assets and incomplete data.
- The `temporal:window` tag (added in Wave 0 structural audit) reflects carrier mortality: the median age of the U.S. water workforce is 48, and the EPA found in 2020 that approximately one-third of the drinking water and wastewater workforce will become eligible to retire over the decade. Their institutional knowledge of buried infrastructure — pipe locations, condition histories, network behavior — is embodied in retiring workers and paper records, and will be permanently lost. This is a closing window for knowledge capture.
- The workforce retirement wave makes this problem time-sensitive.
- Some cities publish water main break data as open data, creating an opportunity for external analysis.
Reconciliation 2026-09-30: Fact repair after reading the whole write-up against every source it names. All four were fetched on 2026-09-30. The McKinsey page could not be reached directly (the connection failed), so the Internet Archive copy was read in full (web.archive.org/web/20260412113643/www.mckinsey.com/industries/infrastructure/our-insights/the-infrastructure-moment); it is the web summary of a 56-page report, and the report PDF was also read and searched, from the Internet Archive copy (web.archive.org/web/20260516105718/www.mckinsey.com/~/media/mckinsey/industries/infrastructure/our%20insights/the%20infrastructure%20moment/the-infrastructure-moment-investing-in-the-expanding-foundations-of-modern-society.pdf). The ASCE 2025 Report Card drinking-water page was read in full from https://infrastructurereportcard.org/cat-item/drinking-water-infrastructure/, and the full national Report Card PDF linked from that page was searched as well. The WEF Global Risks Report 2025 PDF was read in full from https://reports.weforum.org/docs/WEF_Global_Risks_Report_2025.pdf and searched; it says nothing about water pipes, Houston or Atlanta, and no claim in the write-up rests on it. The MDPI paper could not be fetched from mdpi.com (403), so the Internet Archive copy of the article page was read in full (web.archive.org/web/20260608175649/www.mdpi.com/2624-6511/8/4/122), with the citation record from https://api.crossref.org/works/10.3390/smartcities8040122. (1) The text said Houston lost 32 billion gallons of drinking water to leaks in a single year and that Atlanta had over 176 pipe breaks in nine months in 2024 with a 72-hour state of emergency. Neither figure is in any of the four sources (the MDPI paper uses Atlanta only as a water-audit case study). Both sentences were removed, along with the Source Notes bullet that said the Atlanta crisis is well documented in local reporting, which named no source. In their place the paragraph now gives ASCE's figure of approximately 240,000 water main breaks per year costing roughly $2.6 billion in repair and maintenance. (2) The $625 billion figure is now stated as ASCE gives it: EPA's 2023 needs assessment, $625 billion over the next 20 years to reach a state of good repair. (3) The text said the median age of water utility workers is nearly 50 and more than half will approach retirement in the next decade, in both the body and the temporal:window note. ASCE says the median age of the water workforce is 48 and that EPA found in 2020 that approximately one-third of the drinking water and wastewater workforce will become eligible to retire over the decade; both passages now say that. (4) The claims that aging water infrastructure affects every developed country and that developing countries are building systems that will face the same challenges are in none of the sources; the sentence now gives the MDPI paper's figures (average water losses in Europe of about 26%, with Belgium, Austria, the Netherlands, Denmark, Germany and Luxembourg, and separately Australia, listed as losing less than the U.S.). (5) The claim that replacement strategies are often politically driven is in none of the sources; the sentence now gives what the sources say: ASCE's statements that predictive tools are reducing reactive replacements and that only about 30% of utilities have fully implemented an asset management plan, and the MDPI paper's call for proactive rather than reactive approaches. The closing sentence that limited budgets are spent on pipes that just happened to fail was removed because no source says it. (6) The claims that statistical failure models date from the 1990s, consistently underperform, and miss traffic loading, pressure transients and tree roots are in none of the sources and were removed; the passage now gives what ASCE says: many utilities are adopting asset failure prediction technologies, only about 30% have fully implemented an asset management plan, and ductile iron pipes in highly corrosive soil break at six times the rate of those in less corrosive soil. (7) Acoustic leak detection was described as sending sound waves through pipes; the MDPI paper describes noise loggers and listening sticks that use the acoustic signature of the leak, the listening sticks being handheld devices used by highly trained technicians, and the sentence now says that; the phrase about walking every segment was removed because the source does not say it. (8) The text said the 2025 Atlanta study found satellite detection 50% less cost-effective than acoustic. The paper says both were economically favorable over three years and that acoustic generated USD 2.4 million in net benefits, surpassing satellite by 50%; the sentence now says that, and adds the paper's description of satellite detection (synthetic aperture radar, subsurface moisture anomalies, a shorter length of pipe to inspect physically). (9) The sentences on in-pipe inspection robots (SmartBall, PipeDiver) and on digital twins are in none of the sources and were removed; the asset-data point they made is covered by ASCE's 30% asset-management figure given earlier in the paragraph. (10) The Source Notes claim that open main-break data did not exist five years ago is in none of the sources and was removed. (11) The Source line called the McKinsey report a McKinsey Global Institute publication; the page gives it as a McKinsey Report by Alastair Green, Ishaan Nangia and Nicola Sandri, dated September 9, 2025, and the line now copies that. The MDPI title was missing its subtitle and authors; both are now copied from the Crossref record. The WEF and ASCE entries were given their full names, and the ASCE page address that the numbers above come from was added to the Source line. Confirmed in the sources and left as written: more than 2 million miles of pipe (ASCE); pipes mainly built after World War II and reaching the end of their design life (ASCE: post-World War II pipes have an average lifespan of 75 to 100 years and many are reaching or have reached the end of their design life; MDPI: constructed mainly during the 1970s); $625 billion from EPA's 2023 needs survey (ASCE cites the Drinking Water Infrastructure Needs Survey and Assessment, Seventh Report to Congress, 2023); McKinsey's $6 trillion for waste and water infrastructure through 2040; the MDPI study's use of satellite imagery and machine learning; the publication details Smart Cities 8(4):122, 2025. Not checked, because no source names them: the H1's "millions of buried water pipes" (the sources count miles, not pipes); the proposal section and entry points, which are the write-up's own suggestions; the statement that many cities publish main-break reports as open data. No title changed. Genome tags untouched. A second reader went over these repairs against the same sources on the same day; none was undone, and the wording of items 1, 4, 5, 6, 7 and 9 was tightened so that each sentence says no more than its source.