Flint uses predictions to find hazardous water pipes

ActiveRemediation. [1]

XGBoost classification trees with a hierarchical Bayesian model for spatial patterns. An active-learning policy selected further inspections. [1]

The paper names the algorithms but does not identify an XGBoost release or a reproducible deployment configuration. [1]

A research team from the University of Michigan, Georgia Tech, and Brigham Young University. [1]

The system estimated pipe materials and ranked homes for further investigation. [1]

A shared database and a password-protected Flask application let city staff and contractors review maps and add excavation results. [1]

Parcel records, property characteristics, and confirmed pipe materials produced address-level probabilities and inspection priorities. [1]

The system recommended priorities. City officials and contractors decided where to excavate and replaced the pipes. [1]

Crews inspected pipes and recorded their materials. Researchers used those results to update predictions and recommend further inspections. [1]

City officials set priorities, contractors confirmed pipe materials, and researchers maintained the model. [1]

The system used machine learning to predict pipe materials and guide inspections. [1]

Predictions informed inspection and replacement priorities. The paper did not randomize the city’s decisions. [1]

Not assessed

The model helped direct inspections toward homes likely to have hazardous pipes. [1]

Improved accuracy in locating hazardous service lines. [1]

The system was used in the field. Accuracy was tested retrospectively, and monetary savings were modeled. [1]

Residents and the city’s water-infrastructure program. [1]

Residents whose hazardous service lines were identified and replaced. [1]

By September 2017, the program had replaced hazardous lines at 4,419 homes and identified pipe materials at 6,506. [1]

Predictions varied by address and neighborhood. The benefit depended on whether officials used them to set priorities. [1]

Replacement permanently changed the identified service lines. The paper did not measure residents’ later health. [1]

Finding and removing hazardous pipes was the intended benefit. [1]

Residents and officials also sought broad coverage and certainty about individual homes. These priorities could change the order of excavation. [2]

AIBD-0166-E01, AIBD-0166-E02 [1]

Researchers tested predictions on held-out homes and compared inspection strategies in simulations. [1]

Excavation results supplied the pipe-material labels. Validation and simulated work sequences used those records. [1]

The database covered 55,893 parcels. The main accuracy test held out 1,606 homes with known pipe materials. [1]

Pipe-material prediction, inspection success, and modeled cost per replacement. [1]

The cost simulation compared the proposed strategy with the sequence of work in the city’s records. [1]

The held-out prediction accuracy was 91.6%. The simulated strategy lowered estimated cost per replacement by 10.7%. [1]

The estimated cost improvement came from a simulation. The field results do not isolate the model from crews, funding, and policy decisions. [1]

Known pipe materials came from a selected set of inspected homes. The simulated costs omitted travel and other logistical constraints. [1]

Predictions were updated as crews recorded new inspections and replacements. [1]

The paper describes the algorithms and simulations. It does not supply all settings needed to reproduce the deployed system. [1]

Modeled cost savings. [1]

The simulation estimated about US$622 less per replacement, or US$11.18 million if applied to 18,000 replacements. [1]

U.S. dollars, using the unit costs assumed in the study. A separate inflation base year is not stated. [1]

Per replacement, with an extrapolation to 18,000 replacements. [1]

The savings were simulated. They are not an audited reduction in city expenditure. [1]

The modeled savings would reduce the cost of the publicly funded replacement program. [1]

The estimate includes excavation and replacement costs, but not every cost of operating the program. [1]

The simulation assumed US$250 for a hydrovac inspection, US$2,500 for an unnecessary excavation, and US$5,000 for a needed replacement. [1]

Estimated average cost was US$5,818 under the recorded work sequence and US$5,196 under the proposed strategy. [1]

Researchers simulated inspection and replacement sequences, applied fixed unit costs, and extrapolated the difference. [1]

No complete return-on-investment or payback calculation is reported. [1]

University researchers developed the system. Flint officials managed the program, and contractors inspected and replaced service lines. [1]

Flint, Michigan, United States. [1]

The research team developed the models. The city controlled the replacement program and its contracts. [1]

The system depended on digitized parcel records, reliable excavation labels, a shared database, and crews who could inspect and replace pipes. [1]

The research institutions and deployment were in the United States. No international data transfer is documented. [1]

Not assessed

Useful parcel records, verified inspection results, and a working relationship with the people who choose excavation sites are necessary. [1]

Not assessed

A March 2019 court order required the city to prioritize addresses identified by the predictive model, subject to defined exceptions. [3]

Not assessed

The federal court amended the settlement governing Flint’s service-line replacement program. [3]

The parties agreed that the city should first excavate homes with the highest predicted likelihood of hazardous lines. [3]

The city had to prioritize the model’s address lists, document exceptions, and report completed work. [3]

The city had to excavate the listed addresses and replace hazardous lines by 1 January 2020. [3]

AIBD-0166-source-0001, AIBD-0166-source-0002, AIBD-0166-source-0003, AIBD-0166-source-0004 [1] [2] [3] [4]

The original paper establishes the methods and modeled savings. Independent interviews describe implementation. The court order establishes the required prioritization. [1] [2] [3]

English. [1] [2] [3]

The researchers evaluated their own system. The Atlantic adds independent reporting. The court order is evidence of a decision, not an independent estimate of lives or money saved. [1] [2] [3]

The original model paper supplies the validation and simulation. The court order follows the parties’ February 2019 stipulation. [1] [2] [3] [4]

The revised original paper, the complete news article, and the relevant court-order provisions were read. Earlier versions and the underlying stipulation are preserved in the source chain. [1] [2] [3]

Predictions can help locate hazardous service lines and avoid some unnecessary excavations. [1]

The sources are linked to the findings that they support. [1] [2] [3] [4]

The developers reported validation and simulation results. Independent reporting and a court order document how officials used the system. [1] [2] [3]

Not assessed

Identification and replacement of hazardous service lines. [1]

Personal: 2; Organizational: 2; Societal: 3 [1]

Residents benefited from identification and replacement of hazardous service lines, without a measured health effect attributable to the model. The program gained a measured capability to predict pipe materials and prioritize work. Flint replaced hazardous service lines at thousands of homes, with predictions helping prioritize inspections. [1]

The system was used in the field. Accuracy was tested retrospectively, and monetary savings were modeled. [1]

Not assessed

Not assessed

Not assessed

Independently reported [1] [2] [3] [4]

An operational application with retrospective validation and simulation. [1]

Not assessed

Water infrastructure and public health. [1]

The system estimated which homes were most likely to have hazardous service lines. [1]

Field work began in summer 2016. The paper reports replacements and inspections through September 2017. [1]

Researchers worked with Flint’s replacement program from summer 2016. [1]

SHA-256: ded7efa1cccb…

Not assessed

Not assessed

Not assessed

7 revisions. Latest: 25 October 2026.