Industries / Energy, water and waste / Environmental services / Environmental remediation
Pumping and treating groundwater
This work pumps polluted groundwater up through wells and cleans it at the surface, often mainly to stop a plume, the spreading underground body of polluted water, from moving further.
6
Milestones achieved, from 1982 to 2025.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1982
Pumping to hold a plume
In late 1981 Fairchild Semiconductor found that an underground solvent tank at its South San Jose, California, plant had failed, releasing an estimated 60,000 gallons of industrial solvents, including TCA and PCE, into soil and groundwater on and off the site. In 1982 it removed the tank, dug out 3,400 cubic yards of soil and installed a hydraulic control system: wells that pumped groundwater out to hold the polluted water in place, and a plant that treated it.
LessonIn 1986 Fairchild added a slurry wall, a buried barrier of clay-like bentonite, around the site and pumped from inside it so groundwater flowed inward; EPA's report says the wall and the pumping helped speed the cleanup of both soil and groundwater.
Credit
Fairchild Semiconductor; account in EPA's 1995 cost and performance report on the site
1991
Chromium water sprayed on crops
From 1952 to 1964 PG&E's natural gas compressor station at Hinkley, California, sent cooling-tower water holding hexavalent chromium, a toxic form of the metal used to stop corrosion, to unlined ponds, and it seeped into groundwater about 80 feet down. PG&E reported chromium above the state drinking-water standard in late 1987, and the Lahontan regional water board ordered a cleanup on 29 December 1987. Since 1991 the board has let PG&E pump the polluted groundwater and spray it on crop fields, where the soil turns it into trivalent chromium, a less toxic form; a 2014 permit allows up to 500 acres.
LessonA 2015 order set goals for 2025 and 2032, and in July 2026 the board still ordered a study, due in July 2027, to set the final cleanup level for a plume it describes as at least eight miles long.
Credit
Lahontan Regional Water Quality Control Board; PG&E
2004
Computers beat trial and error
The Defense Department's Environmental Security Technology Certification Program (ESTCP) and EPA tested whether computer optimization could design better pumping schemes at three real sites: Umatilla Chemical Depot in Oregon, Tooele Army Depot in Utah and the former Blaine Naval Ammunition Depot in Nebraska. Two university teams ran optimization codes coupled to groundwater flow and transport models, and the firm GeoTrans used the usual trial and error as the control. In every case the optimizers found better solutions, by 5% to 50% and typically about 20%, and the model files, problem statements and results were posted online.
LessonThe report says its second aim, giving each installation pumping plans it could feasibly put in place, was "partially but not fully met.".
Credit
ESTCP; US EPA; teams from the University of Alabama and Utah State University; GeoTrans
2013
126,000 sites still unfinished
A National Research Council committee reported that at least 126,000 US sites with leftover groundwater pollution had been documented, a count it called an underestimate, and put the "cost to complete" at $110 billion to $127 billion. It judged that about 12,000 of them are complex, and that reaching drinking-water standards throughout the aquifer at most complex sites is unlikely within 50 to 100 years.
LessonThe committee warned that the cost figure "is likely to be an underestimate of future liabilities.".
Credit
National Research Council committee chaired by Michael C. Kavanaugh
2019
Reusable resin for PFAS
Ion exchange resin is a bed of plastic beads that grabs charged chemicals, PFAS among them, out of water, and most PFAS resin is used once and thrown away. At the former Pease Air Force Base in New Hampshire, a 200-gallon-a-minute system that washes the PFAS off its resin with a salt and alcohol solution and reuses the resin had treated more than 9 million gallons of groundwater, with all 13 monitored PFAS below detection in the treated water, its operator Wood reported in June 2019 in a paper titled as the world's first such system. In northern Australia a similar plant running since March 2019 had treated about 1.8 billion liters by February 2024 over 26 regenerations with no loss of performance, its authors reported.
LessonWashing the resin moves the PFAS into a small volume of spent liquid that still has to be destroyed; a Defense-funded pilot at Pease in 2022 added a plasma reactor for that step and found the resin still removed more than 95% of PFAS after five regenerations.
Credit
Wood and the US Air Force Civil Engineer Center (Pease); David Kempisty, Steve Woodard and co-authors (Australia)
2025
Pumping falls from favor
EPA's Superfund Remedy Report shows that pump-and-treat was chosen in 85% of Superfund groundwater decisions at its 1991 peak. The share fell below 20% by 2010 and has averaged 23% a year since. In fiscal 2021 to 2023 pump-and-treat was chosen in 22 of 112 groundwater decisions (20%), against 43% for treating the groundwater in place underground.
Credit
US EPA Office of Superfund Remediation and Technology Innovation
Watch for
Reusable resin pays back
In a Defense-funded field test at the former Naval Air Station Joint Reserve Base Willow Grove, Onterris treated about 2.5 million gallons of pumped groundwater to EPA's drinking-water limits for PFAS with a mobile resin system, washing and reusing the resin over three cycles. It reports that washing recovered all the PFAS the resin caught, and that reusable resin breaks even with single-use resin in as little as four years for a 2,000-gallon-a-minute system and 11 years for a 200-gallon-a-minute one. A 2023 Defense-funded study led by Timothy Strathmann of the Colorado School of Mines found that, in its baseline case, single-use PFAS-selective resin had lower costs and environmental impacts than reusable resin or activated carbon.
Achieved whenthe actual operating costs of a full-scale reusable-resin system, checked by an outside party, show it cheaper over its life than single-use resin at the same site.
Claimed by
Onterris (principal investigator Erika Houtz), in a November 2025 fact sheet and a February 2026 final report
Open
01
Rerun the 2004 pumping contest
Port the three 2004 pump-and-treat optimization problems (Umatilla, Tooele and Blaine) from their original MODFLOW96 and MT3DMS files to current open tools, MODFLOW 6 and FloPy, with a scoring script that checks every limit on pumping and concentration, and publish them as an open benchmark. Then test modern optimizers and an AI coding agent against the best 2004 answers.
The briefCredit · unclaimed02
Open PFAS breakthrough records
An open dataset of how fast PFAS break through the carbon and resin beds of real pump-and-treat plants, meaning the point when PFAS start to pass through a bed. Each record gives the vessel, water treated, bed volumes (the number of times the bed's own volume of water has passed through it), PFAS by compound before and after the vessel, and change-out dates, read from public operating reports filed with regulators. Then test whether a chosen prediction method forecasts the change-outs.
The briefCredit · unclaimed
Open 01 · the brief
Rerun the 2004 pumping contest
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
Port the three 2004 pump-and-treat optimization problems (Umatilla, Tooele and Blaine) from their original MODFLOW96 and MT3DMS files to current open tools, MODFLOW 6 and FloPy, with a scoring script that checks every limit on pumping and concentration, and publish them as an open benchmark. Then test modern optimizers and an AI coding agent against the best 2004 answers.
Why it is still open
The model files, problem statements and results have been online since 2004 "for download by interested parties," but in 1990s versions of the models. Crossref lists the project's technical report volumes as cited one to four times, and no later study re-solving the three problems with modern methods was found.
Built with what exists
The ESTCP files on frtr.gov; MODFLOW 6 and FloPy from the US Geological Survey; open optimizers such as SciPy, pymoo or Nevergrad; a coding agent to convert the files and write tests, such as one built on Inkling 2 Small (see the Note "Checked without being asked").
First step
Download the Umatilla files from frtr.gov, run the original model with the 2004 trial-and-error and optimized well settings, and record the objective values.
Done looks like
All three sites run in MODFLOW 6 with heads and concentrations within 1% of the original models at chosen check wells; the 2004 best solutions reproduced within 1% of their reported objective values; and at least two modern optimizers scored on every problem, with code and results public.
Open 02 · the brief
Open PFAS breakthrough records
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An open dataset of how fast PFAS break through the carbon and resin beds of real pump-and-treat plants, meaning the point when PFAS start to pass through a bed. Each record gives the vessel, water treated, bed volumes (the number of times the bed's own volume of water has passed through it), PFAS by compound before and after the vessel, and change-out dates, read from public operating reports filed with regulators. Then test whether a chosen prediction method forecasts the change-outs.
Why it is still open
The authors of a five-year study of a reusable-resin plant in Australia wrote in 2024 that "data and optimisation efforts from full-scale operations has not been widely shared." A 441-day pilot that compared carbon and resin on Minnesota groundwater (Chow and co-authors, 2022) made its data available only on request. EPA publishes open, MIT-licensed models that predict breakthrough in carbon and resin beds, but no open set of field records to test them against was found.
Built with what exists
Public quarterly operation reports, such as those in New York's DECinfo document library (one for a PFAS plant at Stewart Air National Guard Base, which treats stormwater, lists weekly results after each vessel); a model built for extracting from long documents, such as Muse Ember (see the Note "The jobs agents repeat"); an agent that cites the page behind each number (see the Note "Findings that cite notes"); EPA's open carbon model (PSDM) and ion exchange model, as the prediction methods to test; Python.
First step
Pull the last eight quarterly reports for one groundwater plant from a state document library and tabulate one treatment train's results by hand as a test set.
Done looks like
At least 20 vessels from at least five groundwater plants, each with at least a year of data and every number linked to a report page; two people agreeing on at least 95% of fields in a 10% sample; and a stated error, in bed volumes, for EPA's models' forecasts of change-out.