How does AI optimize oil and gas production?

The United States is now the world's largest producer of petroleum and natural gas. Nearly a million wells across the country produce roughly 11 million barrels of crude oil and 4.3 million barrels of natural gas per day. While new technologies have improved oil and gas extraction, the big question for the industry is how to reliably predict well production. To help solve this puzzle, researchers like Dr. Chung-Yan Shih, a senior data scientist at Leidos, are turning to AI. His research for the DOE's National Energy Technology Laboratory (NETL) has helped reveal how machine learning can offer answers to the oil and gas industry as they strive to make their wells more efficient and keep up with demand. 

Q: Why is it so important for oil and gas companies to forecast well production, and what challenges do they face?

Chung: Productivity prediction is critically vital for oil and gas companies because it drives business decisions and helps the industry plan for the future. But, since resources are stored underground, it's been hard to get accurate predictions, and even the best estimates come with significant unknowns. Predicting how much a well can produce throughout its life is referred to as the estimated ultimate recovery (EUR). Initial productions (e.g., first 12-month production) are usually used as proxies to indicate the overall performance of a well since a well often declines fast after its initial peak production. In the long run, many factors may affect the EUR's accuracy. Each well is unique, and the relationships of parameters affecting well production and capacity are complex.

For example, wells differ significantly based on their location, the time they have been in production, their design, and the resource extraction technology that is used, to name a few. However, it's even trickier than that. The reservoir that a well taps into may vary in geologic properties, such as thickness, thermal maturity, or gamma ray levels. Two wells with the same design at different locations are likely to perform differently. As a result, it has been a challenge to precisely predict EUR and to understand the relationships between factors and production.
 

Q: How does AI help us understand these relationships?

Chung: AI is well-suited to help make sense of complicated data sets. One of the key challenges for oil and gas companies is to identify the parameters that drive production. As you can tell from my previous responses, these can be fairly complex. Traditionally, production decline curves and type curves are used to estimate the EUR. Scientists and engineers history-matched curves to well production and averaged the selected curves as type curves to represent the shale play. The process relies heavily on the experiences of the scientists and engineers. However, AI can help to extract production patterns from production data to identify and understand parameter relationships. NETL has highly skilled geologists who estimate geologic properties, and sufficient data for production and well completion, so when we were first asked to help with this problem, we saw great potential in machine learning applications. 

Our goal was to figure out patterns from that data and identify key drivers of production. We tested nine regression machine learning algorithms from simple linear regression to more complicated neural network algorithms. We used these algorithms to predict the first 12-month production of a well, which is a common proxy of EUR. These algorithms took variables from well design and geology, such as a well's lateral perforated length, the gamma ray value of the rock formation, and the amount of proppant, which is sand or a man-made material used in the hydraulic fracturing process. From there, we were able to determine different factors that drive production.

The first study provided great results, but we wanted to take it further because there will always be a gap between the production from the first 12 months of a well's life and the final EUR. And that is what we did in a second study. We used deep learning—a subset of machine learning based on neural networks—to understand a well's production changes over time. We used the same concept of adopting both well completion design and geology parameters to predict the entire production profile of a well, which we used to calculate EUR. Thanks to the capability of deep learning to extract knowledge from complex relationships, we can predict the performance of a well even before it is drilled. 

Q: What was the most important benefit to your customer?

Chung: DOE recently launched an Office of Artificial Intelligence and Technology to “transform DOE into a world-leading AI enterprise.” NETL is interested in determining if machine learning is useful and beneficial to subsurface analysis. The goal with the first study was to understand how the oil and gas industry could better use technology to efficiently extract subterranean resources. The results showed how machine learning could be applied to using existing data to determine production key drivers. AI now appears to provide a way to estimate production for future well development. Through the first and second studies, we've also helped NETL identify potential R&D projects to focus on to help improve resource recovery, such as well stage and spacing optimization, or the impact of different compositions of fracture fluid.  

Q: How does this work impact the future of the oil and gas industry?

Chung: The industry wants to run wells and extract resources as efficiently and cost effectively as possible. We've shown that with AI, they can estimate well production performance based on the well completion design and geologic properties at the well location. Combining the experiences of geologists, engineers, and drillers, they can have a strategic plan for well deployment and optimization to maximize recovery and revenue. Machine learning has also helped us identify and extract complex interactions and relationships from numerous parameters, which the industry can use to design better wells.

Q: What other industries might benefit from this technology?

Chung: The beauty of AI is that it can be applied to many other industries to solve their complex problems. AI is a data-driven approach that is capable of learning complex patterns in the data. Often, the information provided by AI validates the assumptions and reasoning of subject matter experts (such as reservoir engineers and geologists). Sometimes, AI provides a fresh way of thinking and tackling a problem. While those results might be surprising at first, they usually offer valuable insights to help us understand more about the problem. For us, it is always good to work with the experts in the field to make sure AI is learning the tasks we would like it to learn. There have been cases where AI “cheated” in learning in an ingenious way that might be hard for a human to detect. Data scientists and domain experts working together can ask the right questions to learn about a client's goals and fill in the data gaps. These studies have broken new ground, and they show that the types of problems that AI and machine learning can solve are virtually limitless.