Unsupervised Facies Pattern Recognition of Brazilian Pre-Salt Carbonate Borehole Images

Paper Ref: SPWLA-2022-00129
Presenter: Laura Lima Angelo dos Santos
Authors:   Laura Lima Angelo dos Santos, Nadege Bize-Forest, Giovanna da Fraga Carneiro, Adna Grazielly Paz de Vasconcelos and Patrick Pereira Machado, Schlumberger

Presenter Bio: Laura Lima is an Interpretation Development Engineer at the Schlumberger Riboud Product Center (SRPC) in France, where she develops integrated solutions for geology and well construction. Prior to joining the SRPC, she was part of the Schlumberger Technology Integration Center in Rio de Janeiro. She applies her background in geology, data science, and software development to the development of algorithms and UI/UX to make borehole data processing and interpretation more efficient and effective. Laura holds an MSc in Engineering with a specialization in Deep Learning for Borehole Image Classification and a BSc in Geology.

Abstract: We apply our novel automated image interpretation workflow to Brazilian pre-salt ultrasonic borehole image data. We obtain an immediate, un-biased classification of the full data, requiring no further input data beyond the borehole image itself. This interactive solution combines statistical and deep learning algorithms for image embedding to provide data-driven, multi-purpose borehole image interpretation. Borehole images are a source of important information for building static reservoir models. Textures observed in these high-resolution well logs are the results of and provide insights into the different processes that have occurred: from the moment of the deposition until the image acquisition. Each field, reservoir, well, and interval has a unique textural assemblage, consequence of its own depositional facies, diagenetic processes, geomechanics and wellbore conditions or well intervention and completion. Efforts to automate facies interpretation in our industry often rely on applying supervised machine learning models. These supervised algorithms are restricted to executing very specific tasks, based on extensive amounts of consistently labeled data. (Click on link above for full abstract)