Deep learning models could improve oil extraction

Machine Learning

This article has been reviewed in accordance with Science X’s editorial processes and policies. Editors emphasized the following attributes while ensuring content authenticity:


authoritative source


SSL with the Barlow Twins approach to oil and gas data. credit: IEEE Earth Science and Remote Sensing Letter

The Scortec research team has introduced a model that accelerates the planning stage of well development. This helps in getting important data about the well. For example, the model can compare future wells with those already functioning nearby to predict their oil production characteristics and improve well drilling. The research came out IEEE Earth Science and Remote Sensing Letter.

Oil and gas well development can be divided into three phases: field discovery, evaluation, and development itself. Valuations include, for example, oil reserves and their distribution. At this stage, exploration wells are drilled to record indicators such as stratum radioactivity and groundwater fluidity. Later, this information will be useful in making well development decisions.

“Right now, oil field assessments generate a lot of fragmented data, but no one knows how to use that data. The goal is to build a model that creates a vector that fully describes it.” Alexander Marusov, lead author and research engineer at Skoltec.

The vectors returned by the model contain useful information about the wells in compressed form. This model not only predicts its properties, but also helps to deal with the problem of drilling in the wrong direction. When moving into deeper layers, it is important to limit drilling to the same type of rock. If not, you will have to re-drill in the other direction, which is very costly.

“Our model helps us identify rock type and tune drilling. Our model predicts rock type with 82% accuracy, whereas previously the best result was 59%. Our model facilitates decision-making in well development,” said Marusov. I add.

The model was trained through self-supervised learning. This differs from traditional machine learning techniques that require labeled data. They are not used in self-supervised learning. For example, the rover can record radiation and other geophysical signals in exploration wells. Self-supervised learning makes use of these raw data unlabeled.

“Self-supervised learning methods can be divided into contrastive learning methods and non-contrastive learning methods. signal spacing, etc.,” the researchers clarified.

For more information:
Alexander E. Marusov et al., Asymmetric Representation Learning of Intervals from Well Logs, IEEE Earth Science and Remote Sensing Letter (2023). DOI: 10.1109/LGRS.2023.3277214

Source link

Leave a Reply

Your email address will not be published. Required fields are marked *