By Karthika Swamy Cohen
Due to the increasing prevalence of horizontal drilling and hydraulic fracturing methods, interest in unconventional resource exploration has grown, particularly in North America. But with skyrocketing drilling costs, choosing the right locations for new wells and identifying so called, aptly named “sweet spots” for drilling becomes important. The current industry standard of trial-and-error without sufficient information about a drilling location’s physical characteristics no longer works.
In this scenario, data-driven approaches offer good opportunities to effectively characterize sweet spots. Machine learning techniques are particularly well suited to utilize the excessive amounts of data that the industry has been collecting over the last several decades to make informed decisions. They ptovide promising strategies to address the complex exploration problems that arise from unconventional wells, where the underlying physics is not well known, making physical models uncertain.
At the SIAM Annual Meeting being held in Portland, Ore. this week, Ligang Lu of Shell International Exploration and Production demonstrated the use of machine learning methods for sweet spotting and well completion in oil exploration at a minisymposium titled Machine Learning for Unconventional Resource Exploration and Production.
“When a company buys land for oil exploration, it needs to invest in the most prospective land,” he said. “It's a verified risk decision. But if you wait till you get data to decide where to buy, everyone becomes aware, and the price of the land will skyrocket.”
How can machine learning help make that decision to give one a competitive advantage?
Shell uses a collaborative interactive workflow for unconventional exploration, which helps guide decisions on where to drill and how to complete a well.
The workflow first uses data integration and to formulate a set of predictors from original data. This is followed by predictive modeling, where a model is generated based on the predictors and production data using machine learning algorithms.
Lu and his team applied the workflow to unconventional datasets for sweet spot identification. The data set used is an Eagle Ford data set, which was developed over many years gathering information from multiple production wells. The long production history enables one to ensure that the workflow functions well.
The systematic methodology uses machine learning methods to predict production by using well logs and production data from previous explorations.
Lu and his team conducted a chronological test using data from 7,200 oil producers, each with a production history of at least one year. Using all the available information, they formulated a predictive model for all future wells’ production.
The methodology is then applied to real-world data. A heat map shows that the machine learning model predictions are very close to actual data from the Permian basin. The method is more adept at predicting high and low potential areas than the traditional SoPhie model.
Lu then talked about well completion, which is another important aspect of oil exploration. Optimal completion design is essential for well productivity, which is simultaneously influenced by geological factors and reservoir fluid properties.
Data for this part of the study was obtained from 106 horizontal wells in the beta zone of the Permian basin with 12 months of cumulative oil production. Completed lateral length, proppant injected per stage, fracture stage spacing, and geological features—such as water saturation, total organic carbon, production of oil saturation, and medium porosity—are some of the parameters important for well completion.
Analysis of the relation between production and completion parameters requires one to focus on geological effects. Lu’s team focuses on the learning of the functional relationship between production and completion parameters, while treating the collective influence of geological properties as a clustered random effect.
The researchers then use an integrated approach to select clusters and estimate model parameters simultaneously based on an extension of statistical machine learning methods to encourage homogeneity of geological effects within each cluster.