Machine Learning in Geoscience – October SIG Meeting

Machine Learning for Facies Classification and Seismic Inversion to Improve Predictions

With Daria Lazareva, Technical Advisor Petrophysics at CGG GeoSoftware & Shervin Rasoulzadeh, Senior Technical Advisor Seismic Reservoir Characterization at CGG GeoSoftware

As data becomes more and more abundant, machine learning is rapidly becoming a standard technology in the oil and gas industry. Machine learning drives more effective methods and introduces tools and theories for discovering, modeling and extracting patterns and relationships embedded in large datasets. Companies can determine reservoir properties more accurately and more quickly using a new generation of analytics and prediction techniques from machine learning.

We continue to expand on our machine learning technology. Today, machine learning can address complex petrophysical and reservoir engineering challenges by automating mundane routine tasks such as modeling missing log curves and use data clustering for facies classification essential for seismic reservoir characterization or automatically identifying and flagging poor-quality log curves in a project. Users can also take traditional reservoir characterization approaches a step further by deriving actual reservoir properties by predicting 3D volumes that geophysicists and geologists need.

This two-part presentation focuses on machine learning for petrophysical and geophysical data. The potential for machine learning to improve understanding of wells, reservoir and producing fields is virtually unlimited, and to some extent, it all begins with well log data. For data clustering, we are using environmentally corrected, normalized and depth-shifted data to ensure valid interpretation results. We also discuss leveraging machine learning for seismic reservoir studies using deep neural networks.

Deep neural networks require large amounts of seismic and well data. In geoscience studies, the number of wells that has been acquired in a prospective area can be limiting and acquiring new well data can be expensive.  We show how to augment the lack of well data, generate a multitude of synthetic data and train deep neural networks to improve the reservoir prediction. We do this by the use of Rock Physics Models (RPMs), “what if” scenarios (e.g. model changes in porosity, saturation, and lithology) and detailed statistical analysis based on real measurements.

As companies seek to transform their work practices, processes and workflows to ultimately streamline and improve operations, machine learning strengthens the necessary synergy for multi-disciplinary collaboration.

Kindly sponsored by CGG