Industries / Energy, water and waste / Environmental services / Environmental remediation
Site investigation
Site investigation finds out which chemicals are in the soil and groundwater, where they are and how far they have spread, so that a cleanup can be planned.
6
Milestones achieved, from 1979 to 2024.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1979
Solvents shut two town wells
In May 1979 Massachusetts environmental officials found industrial solvents, including trichloroethene (TCE), at 1 to 400 parts per billion in wells G and H in Woburn, which had supplied 30% of the city's drinking water, and the wells were shut down. For EPA, the US Geological Survey then drilled observation wells, ran seismic surveys and test-pumped the wells in 1985 and 1986 to map which ground and river water fed them.
LessonInvestigation data end up in court: at the 1986 trial over the wells, the judge asked jurors to name the month and year each chemical reached them.
Credit
Massachusetts Department of Environmental Quality Engineering; US EPA; USGS hydrologists Charles Myette, Julio Olimpio and David Johnson
1995
Lasers read soil underground
A cone penetrometer is a steel rod that a truck pushes into the ground, with no drilling. In 1995 an EPA-backed test compared two probes that shine ultraviolet laser light through a window in such a rod and record the glow of fuel in the soil: SCAPS LIF, built for the Army, Navy and Air Force, and the Rapid Optical Screening Tool from Loral Corporation and Dakota Technologies. At a Navy fuel test site in Port Hueneme, California, and a tank farm at Sandia National Laboratories in New Mexico, the SCAPS probe agreed with laboratory analysis of soil samples on whether fuel was present more than 90% of the time.
LessonSome natural minerals glow too, which gave many false positives at the Sandia site, and the rod could not push through gravel and cobbles at either site.
Credit
the Tri-Services SCAPS program and the Navy's NCCOSC RDT&E Division; evaluation by Grace Bujewski and Brian Rutherford of Sandia National Laboratories for EPA's Consortium for Site Characterization Technology
1999
Australia sets investigation levels
On 10 December 1999 Australia's National Environment Protection Council approved a national measure for assessing site contamination, made under matching laws of the Commonwealth and each state and territory. It sets out a general process for assessing a site and investigation levels: concentrations in soil and groundwater above which a site needs a closer look. The latest version, compiled on 16 May 2013, also has screening levels for petroleum hydrocarbons and sections on asbestos in soil and on soil vapor surveys.
Credit
National Environment Protection Council
2005
Phase I becomes a rule
A Phase I environmental site assessment is the check a buyer makes before buying land: a review of records and past uses, interviews and a site visit, usually with no sampling. Under the Superfund law, buyers qualify for some protections from cleanup bills only if they first made "all appropriate inquiries." EPA's rule defining those inquiries was published on 1 November 2005 and took effect on 1 November 2006. It defines who counts as an environmental professional, requires an update of any check done more than a year before purchase, and accepts ASTM International's E1527-05 standard as a way to comply.
LessonEach new ASTM edition needs a new EPA rule, and EPA's fast-track rules to accept the 2013 and 2021 editions both drew adverse comments and were withdrawn before ordinary rules replaced them.
Credit
US EPA; ASTM International
2017
Faked samples at a shipyard
The former Hunters Point Naval Shipyard in San Francisco, a Superfund site, housed a Navy laboratory that decontaminated ships exposed to atomic weapons tests from 1948 to 1969. Reviewing a Navy report on radiation soil data from the contractor Tetra Tech EC, EPA's Superfund division manager wrote in December 2017 that the data showed "a widespread pattern of practices that appear to show deliberate falsification," and EPA recommended resampling 90% and 97% of the soil samples from two parcels covering 40% of the shipyard. The Navy agreed to retest all of the contractor's soil samples from those parcels. Two former Tetra Tech EC managers were sentenced to prison for falsifying records, and in January 2025 the firm agreed to pay $97 million to settle federal claims, without admitting liability.
LessonFormer workers described swapping in soil from cleaner areas, so the record of who took each sample, and where, is part of the evidence.
Credit
EPA (Superfund division manager John Chesnutt); US Navy; former workers who reported the practices
2024
One lab method for PFAS
PFAS are a family of long-lasting man-made chemicals used in firefighting foam and many products. In January 2024 EPA released the final version of Method 1633, which measures 40 PFAS in groundwater, surface water, wastewater, soil, sediment, biosolids, landfill leachate and fish tissue, after four drafts from August 2021 and validation studies in several laboratories led by the Defense Department. EPA calls it the only PFAS method validated in multiple laboratories across all these materials; a corrected version, Method 1633A, followed in December 2024.
LessonPFAS investigations keep widening: by 30 June 2026 the Defense Department, which now calls itself the Department of War, had finished first-round PFAS checks at 709 of 727 sites and sent 590 of them on to full investigation.
Credit
EPA Office of Water; the Defense Department's Strategic Environmental Research and Development Program
Watch for
PFAS tested in the field
A team from the University of Tasmania, the scientific-equipment company Trajan Scientific and Medical and the consultancy ADE Consulting Group drove a portable liquid chromatograph and mass spectrometer, an instrument that separates chemicals and weighs their molecules, more than 3,000 km in a mobile laboratory powered by a petrol generator. They report detection limits of 1.9 to 7.5 micrograms per kilogram for 10 PFAS in soil, 24 soil samples analyzed within 8 hours, and close agreement with a central laboratory's results for samples from Budgewoi, New South Wales (a correlation of r = 0.923, where 1 is a perfect match).
Achieved whenan outside laboratory or regulator publishes split samples from several sites showing that the field results match accredited laboratory results, or consultants use the method routinely to decide where to sample and dig.
Claimed by
Ibraam Mikhail, Brett Paull and co-authors, at the Royal Australian Chemical Institute's National Congress, July 2026
Open
01
Score AI on Phase I
A public test set of Phase I environmental site assessments, each with the public records the assessor used and the "recognized environmental conditions" (RECs, signs that a property is or may be polluted) the assessor found, plus a scored run of an AI agent given the same records.
The briefCredit · unclaimed02
A plume-mapping benchmark
A public benchmark built from real monitoring data at many sites. At each site some wells are held back, each method predicts their concentrations from the other wells, and kriging (the standard statistical way to draw a plume between wells), simpler interpolation and machine-learning models are scored on the same splits.
The briefCredit · unclaimed
Open 01 · the brief
Score AI on Phase I
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test set of Phase I environmental site assessments, each with the public records the assessor used and the "recognized environmental conditions" (RECs, signs that a property is or may be polluted) the assessor found, plus a scored run of an AI agent given the same records.
Why it is still open
In a survey published in February 2026 by LightBox, a property-data company, about half of environmental due-diligence professionals already used AI, only a small fraction saw chatbots as effective for the work, and the "overwhelming concern" was accuracy. LightBox wrote that AI cannot "independently determine recognized environmental conditions." Open-source agents that draft Phase I findings appeared on GitHub in 2026, such as matrix-esa-agent, which sorts each finding into one of the standard's classes of environmental condition, but none found publishes a scored test against professionals' findings.
Built with what exists
Public Phase I reports from city and state brownfield programs; EPA's public facility and cleanup-site databases; the Library of Congress's Sanborn fire-insurance maps and USGS historical topographic maps; a model built for long documents, such as Muse Ember (see the Note "The jobs agents repeat"); Gemini Decision 1 to answer a fixed yes-or-no question for each record (see the Note "A thousand questions, one pass"); an agent that cites the record behind each finding (see the Note "Findings that cite notes"); open agents such as matrix-esa-agent as a baseline.
First step
Collect 20 public Phase I reports with their record appendices and list each report's RECs in a spreadsheet.
Done looks like
At least 50 reports, the agent's list of RECs scored for recall and precision against each report's own conclusions, and two environmental professionals reviewing every disagreement to say which side was right.
Open 02 · the brief
A plume-mapping benchmark
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public benchmark built from real monitoring data at many sites. At each site some wells are held back, each method predicts their concentrations from the other wells, and kriging (the standard statistical way to draw a plume between wells), simpler interpolation and machine-learning models are scored on the same splits.
Why it is still open
Published comparisons each use a few sites or simulated plumes. Norman Jones, Jeffrey Davis and Walid Sabbah compared three methods on four field data sets in 2003 and found kriging most accurate at three of them; a July 2026 study compared kriging and machine learning at one chemical industrial park in Shanxi, China; and a 2025 conference abstract from a Jena team compared spatial random forests with kriging for groundwater nitrate across two regions of Germany, a regional map rather than a plume at one site, and found the forests did not clearly beat kriging. No shared benchmark across many sites was found, although California's GeoTracker database alone holds about 285 million lab results from more than 50,000 sites.
scholarsarchive.byu.edumdpi-res.comapi.crossref.orgmeetingorganizer.copernicus.org
Built with what exists
GeoTracker's public lab results and well locations; Python with PyKrige, scikit-learn and XGBoost; a public leaderboard on GitHub or Hugging Face.
First step
Download GeoTracker data for one county and pick 10 sites with at least 20 wells sampled in the same quarter.
Done looks like
At least 100 sites with fixed train-and-test well splits, baseline scores published for at least three methods, and one outside team submitting a result.