Closed-Loop Autonomous Aerial Survey for Locating Undocumented Orphaned Oil and Gas Wells

1. Problem Statement

The United States is spending $4.677 billion to plug orphaned oil and gas wells and cannot find most of them.

Section 40601 of the Infrastructure Investment and Jobs Act appropriated that sum across three programs: $250 million for federal lands and waters, $4.275 billion for state and private lands, and $150 million for Tribal lands (U.S. Department of the Interior, Orphaned Wells Program FY 2025 Annual Report to Congress, November 2025). Through 30 September 2025 the Orphaned Wells Program Office had obligated roughly $1.8 billion, and states had plugged 10,257 wells.

The constraint is not money and not plugging capacity. It is location. The Interstate Oil and Gas Compact Commission counts 141,959 documented orphaned wells on state and private land, up 54% from 92,203 in 2020, and estimates a further 250,000 to 740,000 undocumented orphaned wells whose positions are unknown (DOI FY2025 Annual Report, citing IOGCC 2024). Lawrence Berkeley National Laboratory independently estimates 310,000 to 800,000 (Ciulla et al., 2024). A well that cannot be located cannot be inspected, contracted, or plugged, and the undocumented inventory runs 1.8 to 5.2 times the documented one.

The wells are not inert while they wait. DOI estimates the combined population emits roughly 63 million grams of methane per hour, equivalent to the annual emissions of more than 3.6 million gasoline-powered passenger cars. Field measurement at Oil Creek State Park in Pennsylvania found 62% of ground-verified wells still unplugged and responsible for the majority of leaking wells, at a mean emission factor of 0.027 ± 0.099 kg/day (Saint-Vincent et al., 2021). Unplugged wells also act as conduits between hydrocarbon reservoirs and shallow groundwater, which is why states treat location as a drinking-water problem as much as a climate one.

The buyers are already writing checks. Twenty-six states received Initial Grants totaling $570 million; Formula Grants awarded more than $520 million in Phase 1 and made up to $826 million available in Phase 2; five states drew $78 million in Matching Grants. Federally, the Fish and Wildlife Service received $64 million to plug 564 wells and the Bureau of Land Management $35 million to plug 93, roughly $113,000 and $376,000 per well on the report's own figures. The statutory scope of these grants explicitly includes identifying and characterizing undocumented orphaned wells, not merely plugging them.

Current practice is a three-stage manual pipeline: generate candidates from maps, records, and imagery; fly a pre-planned magnetometer grid; land and post-process over days; then send a ground crew with a backpack magnetometer to walk each surviving candidate.

2. State of the Art

Three technology families compete for the locating problem, and each has a published ceiling.

Records and imagery mining generates candidates cheaply at national scale. Lawrence Berkeley's U-Net over historical topographic maps produced 1,301 candidates across four counties and more than 40,000 km² (Ciulla et al., 2024), and Los Alamos released the imagery-channel counterpart, 120,948 aerial images at 512 × 512 with segmentation masks from National Agriculture Imagery Program coverage at 30 cm to 1 m resolution (Kim et al., 2024). The ceiling is precision: of those 1,301 candidates, 44 were ultimately confirmed. Candidate generation is not location; it is a long list somebody has to walk.

Aerial magnetometry is the confirmation instrument, and its envelope is established: anomalies from vertical well casing are pronounced above background up to 50 m above ground level, with 40 m optimal (Nikulin and de Smet, 2019). NETL scaled the method across six aeromagnetic surveys in Pennsylvania and Wyoming and found more magnetic points than the state databases recorded at every site (Saint-Vincent et al., 2020). The CATALOG consortium concluded that fixed-wing drones carrying magnetometers are the most cost-effective platform for discovering unknown wells (O'Malley et al., 2024). Magnetometry has two ceilings. The first is physical: it detects steel, and in the oldest fields the casing was pulled for salvage or was wood conductor casing to begin with (Saint-Vincent et al., 2021; Li et al., 2026; Thomas and Wang, 2026). The second is computational, quantified in Section 3.

Methane sensing covers what magnetometry misses, because a leaking well announces itself regardless of casing material. Drone-borne statistical methane anomaly detection has been validated against controlled releases and tested at a three-well field site, and its authors are explicit that source attribution and localization remain future work (Thomas and Wang, 2026). A plume says a well is somewhere upwind; it does not say where to dig.

Commercially the field holds adjacent players, not direct ones, and Section 4 names them. None ships an aircraft that decides anything in flight. The gap is therefore neither a sensing gap nor a modeling gap: both sensors are validated and both detection models are published, and nothing closes the loop between them while the aircraft is still airborne.

3. Foundational Research

Marcato A, Colman R, Milazzo D, Guiltinan EJ, Ma Z, O'Malley D, Viswanathan H, Santos JE (2026). "Synthetic Training Enables Deployment on Raw Drone Data: An Attention-Based Framework for Detecting Orphan Wells." Sensors, 26(9), 2573. DOI: 10.3390/s26092573. Los Alamos built a transformer detector that processes hyper-resolute magnetometer traces without the downsampling conventional pipelines require, using on-the-fly windowing and sinusoidal positional encoders for relative positional awareness along the flight line. On purely synthetic data it holds an F1 score above 90% at flight-line spacings up to 140 m, the dominant cost driver in aerial survey. On real drone data the same model achieves 70% recall. Roughly three of every ten real wells under the aircraft are missed, the miss is a sim-to-real transfer failure rather than a sensor limitation, and it is invisible during flight because inference happens on the ground days later.

Ciulla F, Santos A, Jordan P, Kneafsey T, Biraud S, Varadharajan C (2024). "A Deep Learning Based Framework to Identify Undocumented Orphaned Oil and Gas Wells from Historical Maps: A Case Study for California and Oklahoma." Environmental Science & Technology, 58(50), 22194-22203. DOI: 10.1021/acs.est.4c04413. A U-Net detects the cartographic symbols marking wells on historical topographic maps; any symbol more than 100 m from a documented well becomes a candidate. Across four counties in California and Oklahoma spanning more than 40,000 km², the framework produced 1,301 candidates, of which 29 were confirmed from satellite imagery and 15 by field magnetic survey, at spatial accuracy on the order of 10 m. The 1,301-to-44 ratio is what matters commercially: each candidate is a truck, a crew, and a day.

Saint-Vincent PMB, Sams JI III, Hammack RW, et al. (2020). "Identifying Abandoned Well Sites Using Database Records and Aeromagnetic Surveys." Environmental Science & Technology, 54(13). DOI: 10.1021/acs.est.0c00044. NETL flew six aeromagnetic surveys across Pennsylvania and Wyoming and compared detected magnetic points against database records. At every study site, more magnetic points were detected than the databases recorded. From that divergence the team estimated 395,000 to 466,000 total wells in Pennsylvania and 181,000 to 182,000 in Wyoming, and extrapolated a continental United States average of 6.04 ± 19.97 million wells, of which 1.16 ± 3.84 million are abandoned. The very wide confidence intervals are themselves the finding: the national inventory is unbounded at the upper end.

Saint-Vincent PMB, Sams JI III, et al. (2021). "Historic and modern approaches for discovery of abandoned wells for methane emissions mitigation in Oil Creek State Park, Pennsylvania." Journal of Environmental Management, 280, 111856. DOI: 10.1016/j.jenvman.2020.111856. Airborne magnetic and LiDAR surveys over the Pioneer Run watershed, terrain that is steep, densely vegetated, and effectively closed to ground search. LiDAR identified 290 field locations of which 86% proved to be possible well sites; 62% of ground-verified wells remained unplugged and accounted for the majority of leaking wells; the mean methane emission factor for unplugged wells was 0.027 ± 0.099 kg/day. The paper documents the magnetometry blind spot directly: in the oldest fields the casing had been removed or wood conductor casing installed, so the magnetic channel returns nothing.

Thomas WH, Wang C (2026). "Drone-Based Statistical Detection of Methane Anomalies Around Abandoned Oil and Gas Well Sites." Sensors, 26(7), 2205. DOI: 10.3390/s26072205. The University of Alaska Anchorage developed statistical anomaly detection robust to the noise of field environmental sensor data, validated it against controlled methane releases with known emission points, and tested it at a field site containing three abandoned wells with sparse emission profiles. The method identifies areas with elevated methane signals and reduces ground survey scope, and the paper names source attribution and localization as the development still required. It also supplies a national denominator, approximately 3.9 million documented wells whose recorded locations and plugging status vary widely in accuracy.

Li Z, Hollenbeck D, Wu R, Sherman M, Shao S, Sun X, Hassanalian M (2026). "Multi-Agent Reinforcement Learning for UAV-Based Chemical Plume Source Localization." arXiv:2603.11582 [eess.SY], 12 March 2026. A New Mexico Tech and University of New Mexico team addressed the localization step Thomas and Wang left open. The framework coordinates multiple UAVs through virtual anchor nodes, each sensing gas concentration and wind velocity both onboard and through shared measurements, and identifies the source by analyzing the historical trajectory of anchor node placements within the plume. Benchmarked against fluxotaxis, the standard plume-tracing heuristic, it achieved better localization accuracy and better operational efficiency. It is a simulation study: the control policy exists on paper and has never flown.

4. Competitive Landscape

Direct commercial competition is effectively absent, and the reason is structural.

Zefiro Methane Corp. (Cboe Canada: ZEFI) is the closest integrated player. It plugs orphaned wells and monetizes avoided methane through carbon credits, and assembled locating capability by partnership rather than internal development: a January 2025 letter of intent with Geolabe for AI models over satellite methane imagery, a December 2024 engagement with Keynum for an inventory dashboard, and a provisional patent filed by sponsor entity X Machina Sustainable Technologies for an orphaned well data portal. The stack is satellite-down and records-based, and satellite methane instruments resolve far coarser than a wellbore, which is why field verification survives in Zefiro's workflow exactly as in everyone else's.

SPH Engineering supplies the toolchain: UgCS ground-station software with terrain-following flight planning and magnetometer payload support. That is flight planning, not flight decision-making; the mission is composed before takeoff and executed as composed. Geometrics (MagArrow), Sensys, and GEM Systems supply the magnetometers, including the Sensys R3 flown in published surveys at roughly 40 m, with no autonomy layer.

The most serious competition is not commercial. It is the CATALOG consortium, a $30 million Department of Energy program running Los Alamos, Lawrence Berkeley, NETL, Sandia, and Lawrence Livermore against this exact problem since April 2022. CATALOG produced the best detection models, the best datasets, and the platform recommendation the industry follows. The resource asymmetry should be stated plainly: five national laboratories with $30 million are better resourced than any entrant on detection science. What CATALOG lacks is a commercialization mandate, a manufacturing function, and any obligation to ship a certifiable aircraft.

Why has the space not commoditized? Three reasons, only one technical. BVLOS flight has required case-by-case FAA waivers under Part 107, capping survey economics at the range a single operator can see. The buyer is a grant program with multi-year procurement cycles, deterring entrants wanting faster revenue. And the problem spans geophysics, gas sensing, onboard machine learning, and airworthiness, and few teams hold all four. Commoditization becomes plausible roughly 18 to 30 months after Part 108 takes effect, when conventional aerial contractors can enter.

5. Total Addressable Market

The market is defined narrowly: autonomous aerial survey services and onboard autonomy systems for locating and positionally verifying undocumented orphaned oil and gas wells in the United States and Canada. It is not the drone market, the methane monitoring market, or the well-plugging market.

Bottom-up. Every input is from primary federal reporting or peer-reviewed literature, and every derived figure is labeled as derived. The appropriated pool is $4.677 billion under IIJA Section 40601, of which $1.8 billion was obligated through 30 September 2025, leaving approximately $2.9 billion unobligated. The State Program, holding $4.275 billion, explicitly covers identifying and characterizing undocumented orphaned wells alongside plugging. DOI does not break out the locating share, and this brief does not pretend otherwise; two anchors bound it. First, the Department of Energy has separately committed $45 million to locating and characterization alone: $30 million to CATALOG under IIJA, and up to $15 million announced 13 September 2024 for Undocumented Orphaned Well Characterization and Remediation, structured as up to seven awards of $1.2 million to $3.4 million with 20% cost share. That $45 million is directly addressable technology-development spend today. Second, applying a 3% to 8% locate-and-characterize share to the $4.1 billion in direct State Program funding, a band consistent with site-characterization fractions in comparable environmental remediation programs and stated here as an assumption rather than a source, yields $123 million to $328 million across the program's remaining life, roughly $20 million to $55 million per year through 2031.

Physical demand outlives the appropriation: states plugged 10,257 wells in roughly three years, so clearing even a 250,000-well inventory is a multi-decade program bounded by locating rather than plugging.

Top-down cross-check. The methane emissions monitoring market was valued at $4.8 billion in 2025 and is projected to reach $11.6 billion by 2034, a 10.3% CAGR, with software at approximately 31.5% of 2025 revenue and growing fastest at 13.7% CAGR (Dataintelo, Methane Emissions Monitoring Market Research Report, 2025). Orphaned well location is a small, underserved segment inside that envelope, and software being both highest-margin and fastest-growing is consistent with a bottom-up view placing durable value in the autonomy stack rather than the airframe.

SAM. At TRL 4, with no certified BVLOS mission flown, the addressable slice in the first 36 months is the $45 million in DOE technology-development funding plus early survey contracts in the three to five states with large undocumented inventories and active Formula Grant Phase 2 draws. At a 10% to 15% share of the midpoint services band, SAM is approximately $3 million to $8 million in annual services revenue by year four.

Payment pathway. No CPT or HCPCS analogue applies, because this is not a reimbursed clinical service, and any brief implying otherwise has misread the buyer. Three rails carry the money, two documented in Section 1: DOI Orphaned Wells Program Office grants to states, and direct federal contracting through BLM, BSEE, FWS, USFS, and NPS. The third is the voluntary carbon market via the American Carbon Registry Methodology for the Quantification, Monitoring, Reporting and Verification of Greenhouse Gas Emission Reductions from the Plugging of Orphaned Oil and Gas Wells, published May 2023 as the first of its kind for the United States and Canada. It carries a caveat belonging in any investment memo: ACR made version 1.0 inactive effective 9 May 2025 pending revisions, with already-verified projects and issued credits unaffected. Carbon revenue is real but in methodological transition, and a plan leaning on it as primary revenue is mispricing risk.

6. Research Gap and Commercial Opportunity

Every deployed locating workflow in this field is open loop. The aircraft flies a pre-planned grid, lands, and its data is post-processed days later, after which a ground crew walks the surviving candidates. The aircraft decides nothing while airborne.

That architecture produces three quantified failures. The detector loses roughly 30% of real wells and nobody learns it until after the aircraft has left the county (Marcato et al., 2026). Candidate lists confirm at roughly 3%, 1,301 to 44, so truck rolls consume the budget (Ciulla et al., 2024). And the two modalities that would cover each other's blind spots are never fused in flight, so a well with pulled or wooden casing that is actively venting is invisible to the magnetic channel and unlocalizable by the methane channel (Saint-Vincent et al., 2021; Thomas and Wang, 2026).

The opportunity is the closed loop. A survey aircraft running detection inference onboard, fusing magnetic and methane channels in real time, and re-tasking itself to confirm a candidate before leaving the site converts a three-stage pipeline with multi-day latency into a single sortie returning confirmed coordinates. The enabling control policy is already published and beats fluxotaxis on both accuracy and operational efficiency (Li et al., 2026), and it has only ever run in simulation.

Why have the national laboratories not built it, holding the models, the data, and the platform recommendation? Not capability and not funding. Mandate. CATALOG's charter is to develop technologies and methodologies for finding and characterizing undocumented wells, and it discharges that charter by publishing models, releasing datasets, and issuing platform guidance. It has no product function, no design-for-manufacturability discipline, and no route to airworthiness certification, because none of those is what a national laboratory is instrumented to produce. The Los Alamos detector was evaluated offline on recorded traces, not deployed as an onboard inference target with a latency budget and a power envelope.

The other incumbents have non-technical reasons too. Sensor vendors would be competing with their own integrator customers; flight-software vendors sell pre-flight determinism, so in-flight autonomy is a different product with a different certification burden; and Zefiro's economics make locating a cost line to outsource rather than a capability to own, which three partnerships instead of one internal build confirm.

7. Comparable Funded Projects

Agencies have committed real money to adjacent approaches, validating both the problem and the willingness to fund technology rather than only services.

CATALOG (Consortium Advancing Technology for Assessment of Lost Oil and Gas Wells). Funder: DOE Office of Fossil Energy and Carbon Management, via IIJA. $30 million. Institutions: Los Alamos, Lawrence Berkeley, NETL, Sandia, Lawrence Livermore. Kickoff April 2022, five-year program. Every national-laboratory result cited above traces to it, establishing that DOE treats locating as a fundable research problem in its own right.

Undocumented Orphaned Well Characterization and Remediation. Funder: DOE Office of Fossil Energy and Carbon Management. Up to $15 million, announced 13 September 2024, up to seven awards of $1.2 million to $3.4 million with 20% minimum cost share, applications due 13 November 2024. Focus: advanced borehole remediation, wellbore characterization, long-term monitoring. Eligibility explicitly extends to for-profit organizations, the commercially decisive signal that DOE is buying technology development from companies here.

NSF Award #2323050, "SCC-PG: Sustainable Vertiports for Bringing Autonomous Drone Swarm Inspection to Oil and Gas Industry." PI Sihua Shao, New Mexico Institute of Mining and Technology. $149,999, 1 October 2023 to 30 September 2025, Smart and Connected Communities Planning Grant. The closest funded analogue, and the link is not coincidental: PI Shao co-authored the 2026 plume-localization paper cited above.

NSF Award #2133029, "NRI: An Ecologically Curious Robot for Monitoring Coral Reef Biodiversity." PI Yogesh Girdhar, Woods Hole Oceanographic Institution. $1,499,986, 1 March 2022 to 31 December 2025, National Robotics Initiative. The methodological neighbor: a vehicle that decides in the field what to sample next from what it has already observed, the same informative-path-planning problem in a different medium. It sets the NRI scale for this class of work at roughly $1.5 million over four years.

8. Opportunity Assessment

TRL 4, with a specific evidence chain. Magnetic detection is field-validated at TRL 5 to 6: anomalies confirmed up to 50 m AGL with 40 m optimal, across six NETL surveys in two states plus independent UAV work in New York. Onboard-capable detection models sit at TRL 4: transformer inference validated on real drone traces at 70% recall, offline. Methane anomaly detection sits at TRL 4 to 5, validated against controlled releases at a three-well field site. The autonomous localization policy sits at TRL 3, simulation only. The integrated closed-loop system exists at no readiness level, which is the definition of TRL 4 for the system as a whole; TRL 5 requires a flight-integrated payload demonstrating in-flight re-tasking in a relevant outdoor environment.

Risk 1: onboard inference misses the latency and power budget. A transformer processing hyper-resolute magnetometer traces was designed for offline throughput, not a battery-powered fixed wing at 40 m AGL. Mitigation is quantization and distillation to an embedded accelerator with a hard go/no-go: if the distilled model cannot sustain inference at the sensor's native sample rate inside the payload power envelope while holding recall within 5 percentage points of the offline model, the architecture falls back to onboard anomaly triggering with full inference deferred, which still delivers re-tasking at reduced precision.

Risk 2: 70% real-data recall is a domain-shift problem that closed-loop flight does not automatically fix. If the model misses a well, re-tasking never triggers. Mitigation is to make the methane channel an independent trigger rather than confirmation-only, so a well invisible to magnetics because its casing was pulled can still initiate a localization maneuver. This is why dual modality is architectural rather than incremental.

Risk 3: regulatory timing slips. Autonomous survey economics depend on BVLOS, but every capability is demonstrable within visual line of sight under Part 107, so a slip delays scale-up rather than invalidating the program.

Regulatory pathway. This is not an FDA-regulated system and no 510(k), De Novo, or PMA pathway applies. The governing regime is the Federal Aviation Administration, and operations currently run under 14 CFR Part 107 with case-by-case BVLOS waivers. The relevant change is proposed Part 108: the FAA published the NPRM "Normalizing Unmanned Aircraft Systems Beyond Visual Line of Sight Operations" on 7 August 2025, the comment period closed 6 October 2025 with more than 3,000 comments, and the FAA reopened comment on 28 January 2026 on ADS-B Out equipage, alternate electronic conspicuity devices, and detect-and-avoid requirements. Final rule publication is expected during 2026 with implementation 6 to 12 months after. Secondarily, surveys over federal lands require bureau land-management authorization, and ground-truthing crews fall under OSHA hazardous-atmosphere provisions. Regulatory burden functions here as a moat: Part 108 will impose airworthiness, operator certification, and detect-and-avoid requirements a component vendor or two-person survey shop cannot readily satisfy, and an operator that built its stack against them enters the licensed era with a 2 to 3 year lead.

Locked versus adaptive algorithm. The distinction FDA formalized for medical AI has a direct aviation analogue and is the most consequential design decision here. A model that keeps learning on-aircraft is, from a certification standpoint, a system whose behavior at hour 1,000 is not the behavior certified at hour zero. The recommended architecture is a locked inference model in the flight-critical path, retrained and revalidated on the ground under version control, with adaptation confined to the non-flight-critical survey-planning layer inside pre-declared bounds. This mirrors FDA's Predetermined Change Control Plan framework, and adopting that discipline voluntarily is the fastest route through an aviation certification process that has no equivalent framework of its own yet.

9. Team Requirements

The capability profile this opportunity demands is unusual, and the mismatch between it and the profile of the institutions holding the science is the basis of the opening.

Machine learning and autonomous control systems. Distilling a transformer magnetometer detector to an embedded inference target under hard latency and power constraints; flight-adapting the multi-agent reinforcement learning localization policy from simulation; designing the sensor-fusion layer that arbitrates between magnetic and methane triggers; and building the evaluation methodology that proves closed-loop performance against the open-loop baseline of 70% real-data recall.

Applied physics, geophysics, and sensor systems. Magnetometer payload design and calibration; magnetic anomaly interpretation against the 40 m altitude envelope; methane plume physics and the wind-field estimation the localization policy consumes; and experimental design for controlled-release validation trials.

Manufacturing engineering. Required from the first month, not the last. A payload that works once on a bench is not a product. The deliverable is one built repeatably to tolerance, surviving field handling, carrying the documentation trail an FAA airworthiness process demands. Design-for-manufacturability discipline at the prototype stage separates a demonstration from a deployable system, and its absence is the most common reason funded research in this class stalls between TRL 4 and TRL 7.

This is deliberately a computational and manufacturing profile, not a full-stack geophysical-services profile. An entrant competing head-on with a $30 million five-laboratory consortium on detection science would lose. An entrant supplying the onboard autonomy layer and the manufacturable payload that the consortium's own models require in order to leave the notebook is complementary to it, and positioned to license, partner, or subcontract rather than displace.