Scattered across the United States are the remnants of nearly 170 years of commercial drilling: hundreds of thousands of oil and gas wells that were drilled, abandoned, and simply forgotten. They appear in no formal records. They have no known operators. And in many cases, no one knows exactly where they are.
That invisibility carries real consequences. Unplugged wells can leach oil and chemicals into nearby water sources — and silently emit methane, a greenhouse gas roughly 28 times more potent than carbon dioxide, into the atmosphere. Estimates suggest there may be between 310,000 and 800,000 of these undocumented orphaned wells across the country. The challenge isn’t just plugging them. It’s finding them first.
A hidden legacy of 170 years of drilling
The scale of the problem traces back to the very origins of the American oil industry. Regulations for drilling and plugging emerged inconsistently across states — often decades after the first wells were sunk — and many early boreholes were simply left open or filled with makeshift plugs. That patchwork history means oil, gas, brine, and chemicals can still escape into surrounding soil and water today.
The risks are concrete. Unplugged wells can push toxic substances like benzene and hydrogen sulfide into the air, contaminate nearby water sources, and leak methane — a greenhouse gas with roughly 28 times the heat-trapping potency of carbon dioxide over a hundred-year period, and an even higher warming impact over shorter timeframes.
The 2021 Interstate Oil and Gas Compact Commission estimate of 310,000 to 800,000 undocumented orphaned wells reflects how much uncertainty remains. No one has a precise count, because by definition these wells are absent from formal records and have no known or financially solvent operators. They are, in the most literal sense, lost.
Teaching AI to read maps from another era
Finding them required a creative solution — one that looked backward before it could look forward. Since 2011, the U.S. Geological Survey has digitized roughly 190,000 historical topographic maps covering the period from 1884 to 2006. Each map is georeferenced, meaning every pixel corresponds to real-world coordinates.
Between 1947 and 1992, USGS quadrangle maps used a standardized symbol for oil and gas wells: a hollow black circle. Simple enough for a human to recognize at a glance, but impossible to search manually across tens of thousands of maps at any meaningful scale.
The Berkeley Lab team started with nearly 100 California maps, annotated by hand to build a training dataset. The AI model learned to identify those hollow circles while filtering out visual noise — cul-de-sacs, the number “9,” the letter “o,” and other circular patterns that could generate false positives. Every detected symbol was then cross-referenced against known documented wells, and any symbol located more than 100 meters from a known well was flagged as a potential undocumented orphaned well.
“AI can enhance our understanding of the past by extracting information from historical data on a scale that was unattainable just a few years ago,” said Fabio Ciulla, a postdoctoral fellow at Berkeley Lab and lead author of the study published in Environmental Science & Technology.
1,301 potential lost wells found across four counties
The team applied the algorithm to four counties with significant early oil production histories: Los Angeles and Kern counties in California, and Osage and Oklahoma counties in Oklahoma. Across those four areas alone, 1,301 potential undocumented orphaned wells were identified.
Caution was built into the methodology from the start. The researchers deliberately accepted more false negatives than false positives — meaning they would rather miss a well than wrongly flag a location. That conservative approach suggests the true number of undocumented wells in those counties is likely higher than what the algorithm surfaced.
Field verification put the method to a real test. Researchers carried magnetometers across predicted well sites in grid or spiral patterns, detecting buried metal casings by the way they disturb local magnetic fields. Verified undocumented orphaned wells were found an average of just 10 meters from where the AI had predicted — a margin that points to strong locational accuracy at county scale.
From the ground to the sky: a multi-layer detection toolkit
Confirming wells one by one on foot works, but scaling that process across millions of square miles demands more. CATALOG researchers are building a layered toolkit that combines ground-level sensors with aerial platforms.
Drones preprogrammed with flight routes carry magnetometers suspended on 9-foot cables — the length needed to keep the sensor away from the drone’s own electronics. Separate drones carry methane sensors that factor in concentration, wind speed, and wind direction to pinpoint leak sources. Hyperspectral cameras detect methane plumes in wavelengths invisible to the human eye.
For teams that can’t afford high-end sensors, Berkeley Lab scientist Sebastien Biraud’s group is developing lower-cost alternatives. A setup combining an anemometer, a gas analyzer, GPS, and a small fan can estimate methane leak rates in roughly five minutes. “We need to know if it’s not leaking, if it’s leaking between 10 and 100 grams per hour, or if it’s leaking kilograms per hour,” Biraud said. Spanning nearly 1.5 million acres, the Osage Nation serves as a key proving ground for evaluating equipment performance.
A national program racing to plug the past
All of this work sits within CATALOG — the Consortium Advancing Technology for Assessment of Lost Oil and Gas Wells — led by Los Alamos National Laboratory and involving five national laboratories in total, including Berkeley Lab, Lawrence Livermore, the National Energy Technology Laboratory, and Sandia.
Once a well is located and assessed, the remediation process is straightforward in principle: fill the borehole with cement, cutting off pathways for oil to reach groundwater and methane to reach the atmosphere. New regulations now require methane emissions to be measured both before and after plugging, to confirm the work was effective and to quantify the program’s broader climate impact.
The longer-term goal is making every tool in this toolkit affordable and accessible enough that any state or county program can adopt it. As the AI mapping method is refined and the sensor toolkit expands, each verified location becomes one more well that can be assessed, prioritized, and eventually plugged — turning an invisible hazard into a problem with a known address and a clear path toward resolution.
Carlos is an engineer with strong expertise in technical and industrial topics. He previously worked at international companies such as Siemens and speaks Spanish, German, English, and Italian.
