AI Finds Forgotten Oil Wells

The proliferation of abandoned oil and gas wells across the United States poses a significant environmental concern, with an estimated 300,000 to 800,000 undocumented orphaned wells (UOWs) leaking toxic chemicals and greenhouse gases into the atmosphere.

To combat this issue, researchers from Lawrence Berkeley National Laboratory have developed an artificial intelligence model that can accurately locate these hazardous wells at scale. By training a vision-language model on digitized maps of the US created between 1947 and 1992, the team has successfully identified over 1,300 potential UOWs across four counties in California and Oklahoma, with accuracy rates ranging from 31% to 98%.

This innovative approach has far-reaching implications for environmental remediation efforts, as it enables the efficient location and sealing of these leaky wells, thereby mitigating their harmful impact on the environment.

Introduction to Undocumented Orphaned Wells

The United States is home to an estimated 3.7 million oil and gas wells, with a significant subset of these being undocumented orphaned wells (UOWs). These UOWs are wells that do not appear on official records, have no known owner, and are often left unsealed, potentially leaking toxic chemicals and greenhouse gases into the environment. The lack of documentation and ownership makes it challenging to locate and seal these wells, which is essential for preventing environmental harm.

The Lawrence Berkeley National Laboratory (LBNL) has developed an AI model capable of accurately locating UOWs at scale. This model uses a vision-language approach, trained on digitized maps of the US created between 1947 and 1992. The uniformity and georeferencing of these “quadrangle” maps make them ideal for identifying wellheads and other features. By leveraging this technology, researchers aim to help states identify UOWs and take steps towards mitigating their environmental impact.

The development of this AI model is part of a broader effort to address the issue of UOWs in the US. The Department of Energy has initiated a program to support states in identifying these wells, and the LBNL study is a key component of this initiative. By refining the model and expanding its application to other locations, researchers hope to make a significant contribution to the identification and remediation of UOWs.

Methodology and Model Development

The researchers trained their wellhead-finding model on maps of two California counties, Los Angeles and Kern, which were top oil and gas producing counties in the early 1900s. To ensure accuracy, they manually annotated 79 digitized, georeferenced maps of these counties to identify every wellhead symbol. The model was then fine-tuned using these updated maps, allowing it to learn the patterns and features associated with wellheads.

To identify UOWs, the researchers cross-referenced wellheads identified by their model on historical quadrangle maps with locations in a database of known wellheads in LA and Kern counties. When the model identified a new wellhead that was more than 100 meters from a known wellhead, it was treated as a potential UOW. This approach enabled the researchers to identify 1,301 potential UOWs across four counties in California and Oklahoma.

Model Accuracy and Transferability

The accuracy of the model in identifying UOWs varied, ranging from 31% to 98%. In more rural areas, the model was highly accurate, while in urban areas, it was less accurate due to factors such as paved-over wellheads or confusion with other symbols. Despite these challenges, the model demonstrated transferability, performing similarly well when applied to Oklahoma’s Osage and Oklahoma counties without requiring retraining.

The model’s ability to generalize across different regions is a significant advantage, as it allows researchers to apply the technology to various locations without needing to retrain the model for each area. This feature will be essential for scaling up the identification of UOWs nationwide.

Future Directions and Implications

The LBNL study has made a crucial contribution to the development of AI-powered tools for identifying UOWs. The researchers plan to continue refining their model, expanding its application to other locations, and collaborating with states interested in using their work to identify UOWs. By leveraging this technology, states can take proactive steps towards mitigating the environmental impact of these wells.

The identification and remediation of UOWs have significant implications for environmental protection and public health. By sealing these wells, states can prevent the release of toxic chemicals and greenhouse gases into the environment, reducing the risks associated with groundwater contamination and climate change. The development of AI-powered tools like the LBNL model will play a vital role in addressing this critical issue.

Conclusion

The Lawrence Berkeley National Laboratory’s AI model for identifying undocumented orphaned wells represents a significant breakthrough in the effort to address the environmental impact of these wells. By leveraging digitized maps and vision-language technology, researchers have developed a tool that can accurately locate UOWs at scale. The model’s transferability and ability to generalize across different regions make it an essential component of any nationwide effort to identify and remediate these wells. As the US continues to grapple with the challenges posed by UOWs, the development of AI-powered tools like the LBNL model will be crucial for protecting the environment and public health.