Machine Learning Lithofacies Prediction and Deep Learning Fault Detection Transform Reservoir Characterization – Faster and More Economical vs. Traditional Methods – Webinar Summary/Review. June 2, 2026.

      This webinar was presented by three people from Geophysical Insights. I had some trouble with the audio, so I am relying a lot on the figures shown. This is a software-based analysis of AI/ML geoscience workflows. Al Green, Director of Geophysical Insights, was the first presenter. He gives the AI/ML workflow in terms of four key ideas:

1) Lithofacies prediction with faster integration of machine learning with well logs; 2) Thin bed detection; 3) Apply information theory to identify optimum seismic attributes; 4) Generate fault volumes in 2D and 3D using synthetic fault models.

     I believe the software utilized is called Paradise. He speaks of the workflow as thought-flows – from stratigraphic analysis to lithofacies prediction to fault detection. Seismic and well logs are used for lithofacies prediction. A self-organized map, or SOM topology, is generated by AI/ML, which is converted to lithofacies topology and then to geo-bodies with calculatable volumetrics. The workflow enables a means of interrogating 3D results.

     The next speaker was Alvaro Chaveste, Sr. Geophysical Advisor at Geophysical Insights. The slide below shows seismic, well log, and resulting lithofacies prediction of a case study in the West Desert of Egypt.

     As shown below, the seismic interpretations and well logs are integrated to get the lithofacies prediction.

     SOMs, or self-organized maps, with a clustering of seismic attributes are shown below, followed by how to decide optimum attributes to use and attribute selection. Attributes are selected using Mutual Information, an idea from information theory that measures how much information two random variables share. It determines how much knowing one variable reduces uncertainty about another variable. Variables include porosity, water saturation, etc., and petrophysical properties. The flow is to first generate an SOM, then compare it to seismic.

     Machine learning for lithofacies prediction involves creating a table and histogram showing the matching of predictions from different sources, such as finding hydrocarbon saturation and distinguishing it from water saturation, to get a better SOM that shows the hydrocarbon reservoir. The SOM can now be generated in 3D, and volumetrics can be calculated.

     The next speaker was Fabian Rada, Geophysical Consultant at Geophysical Insights. His talk was about complex clastic gas reservoirs offshore India. The goal is to reduce uncertainty in a complex reservoir. The first slide below shows the geology of the complex stacked reservoirs, and the second one shows the challenge: modeling a thin reservoir with a resolution of 5 meters.

     He goes through the workflow that utilizes deep learning fault detection and moves from ML lithofacies prediction to the generation of ML geo-bodies, in this case, gas sands. The seismic data trains the model. The deep learning workflow for fault detection is shown below.

     The next step is to find the ML-generated faults on seismic. Then they can be integrated and shown on contour maps. Finally, the ML lithofacies prediction can be used for geomodelling, and porosity can be modeled from lithofacies volumes. His conclusions are shown below, and they include that the method can be used to analyze complex reservoirs, including reservoir compartmentalization, which traditionally can be problematic for prediction.

     One question from the Q&A was whether the technique can be used for 2D data sets. The answer is yes, it can, but one may lose the homogeneity of the data.

     This webinar was tough to follow at times, and utilizing the technique requires complex software and large datasets. However, it appears that it can be successfully utilized for geomodelling complex reservoirs.

 

 

 

 

 

 

 

;