BIO: Daniel Austin
Daniel is Global Business Development Manager for Earth Science Analytics and has over

10 years of experience in geosciences. Originally from the United Kingdom, Dan has a range of experience in E&P especially with the Atlantic transform margins and West Africa. He received both his bachelor’s and master’s degrees from the University of Southampton.

SUMMARY:
The landscape of the energy industry is rapidly changing and we are
undergoing unprecedented disruption, challenges and opportunities. As the
energy transition gathers pace many new and innovative projects are exploring
areas where traditional data sets are sparse, inconsistent or absent entirely.
Rock images are one of the most abundant sources of ground truth for
geological interpretation, however historically we have been limited in our
application of this data as part of decision making processes. Recent efforts to
collect and prepare image data in a standardised and consistent format have
opened up new opportunities for analysis driven by machine learning
technologies.
The potential gains from the systematic analysis and management of rock
images is therefore unprecedented, however we are not immune to the
challenges faced by other industries when adopting machine learning
workflows. While one-off projects often result in novel breakthroughs within
their local datasets, they frequently struggle when introduced to new data,
users and use cases.
Here we present how Earth Science Analytics have developed a practical,
elegant, and scalable solution to streamlining rock image interpretation.
Through close collaboration with Rockwash Geodata we provide a toolkit that
allows geoscientists to laser focus on solving specific problems while
maintaining a consistent and accurate framework.
Through combining the best patrices of machine learning operations (ML-Ops)
with critical domain expertise we have achieved high quality results on an
unprecedented scale. and find new applications of image data in areas as
diverse as geothermal, carbon capture and storage and extraterrestrial geology.
We will present a selection of case studies from our recent work in applied
machine learning and discuss some of the future challenges and opportunities
presented by the changing landscape.

Dismiss

BIO: Karim Bondabou
Currently holds the position of Surface Geology Product Champion of Well Construction Measurement Division from Schlumberger located in France. He currently leads the development and introduction of new technologies related to Surface Formation Evaluation on rocks generated while drilling commonly called “cuttings” to improve reservoir characterization real-time. During his 13 years in Surface Logging, Karim was highly involved into development, global deployment, and data interpretation of various measurement technology to extract the maximum value of drilled cuttings at rig-site through various position. He holds a Master degree in Applied Geology with specialization in Geological Reservoir Characterization from the University of Sciences and Technology of Montpellier (France).

SUMMARY:
Rock lithological identification, quantification and description from drill cuttings is still a fundamental task performed by surface logging to aid in reservoir characterization at the rig site. All first insights of physical parameters for drilled geological formation occur in real-time at the rig site and often aid in important well decisions and provide key characteristics of the reservoir model. The analysis and interpretation of drilled cuttings has not evolved in over 65 years and is a tedious human dependent process, which is often error prone. This presentation will describe a new methodology, which utilizes digital images and smart processing based around Artificial Intelligence and Machine Learning algorithms to automate the lithological interpretation and automatically extract rocks features as a direct result of digital measurements. The methodology directly mimics a geologists behavior to classify lithologies, by looking at rock color, grain size, fabric (texture) and mineralogy. This digital solution removes any ambiguity in sample descriptions while enabling remote operations by allowing experts to operate where is matters, reduce the skills needed to describe the samples and at the same time unlock new digital answer products. This new approach will boost the value of drilled cuttings for real-time geological interpretation and improve the flexibility of this data source for sub-surface understanding thanks to Artificial Intelligence and Machine Learning.

Dismiss

BIO: Issac Sujay Anand John Jayachandran
Issac is a Ph.D. candidate at the department of Geology and Geophysics at Texas A&M University, and is currently based in the Qatar branch campus. His research is focused on the intersection of computer vision, artificial intelligence, and geology. In that pursuit, his Ph.D. dissertation ranges from quantifying calcite microcrystals in SEM images to segmenting microfracture and pore systems in thin-section scans. Issac earned his undergraduate and master's degrees in Geology from Khalifa University (formerly The Petroleum Institute), Abu Dhabi. Along with Mokhles Mezghani, Issac serves on the Awards sub-committee for IRIS 2022.

SUMMARY:
The importance of microfractures is being realized owing to their utility in facilitating flow in tight reservoirs and serving as an archive for microtectonics. While solutions to segment macrofractures from pores in images exist given their outsized importance, to our knowledge, there are no solutions that target the segmentation of microfractures from pores in microscopic images. The added complexity is that, as objects, microfractures and pores are typically similar in size, necessitating a shape-based approach. Aspect ratio (the ratio of major axis length to minor axis length) is commonly assumed in geoscientific literature as the only shape feature needed to discriminate the two classes. However, we hypothesize that aspect ratio alone is insufficient, and a multi-dimensional approach using several independent shape features within a supervised machine learning (ML) context is required. To that end, we tested the following hypotheses; first, unsupervised ML, namely Principal Components Analysis (PCA), cannot accurately classify microfractures from pores using simple geometric shape features such as aspect ratio, compactness, and solidity among others. Second, supervised ML methods will be highly accurate in discriminating the two classes using the same features. Third, complex, non-linear supervised ML methods (Quadratic Discriminant Analysis (QDA), Random Forest, and Support Vector Machines (SVM) with radial and polynomial kernels) will significantly outperform simpler, linear models (multiple logistic regression, linear discriminant analysis (LDA), and linear SVM). Lastly, the aspect ratio is not the most important differentiating shape feature between the two classes. The hypotheses were tested on 18 carbonate thin sections, consisting of natural and artificial microfracture, sourced from a range of subsurface and outcrop sites in the USA. The thin sections were scanned whole using a Nikon CoolScan 8000 film scanner with a resolution of 6.35 microns/pixel. As anticipated, PCA performed poorly compared to supervised machine learning methods which exhibited testing accuracies within 85-90%. We also report that, rather surprisingly, the more complex non-linear supervised models did not significantly outperform the simpler linear models, thus facilitating the use of the simpler models to solve this challenge. Finally, we report that while aspect ratio ranked highly across all supervised ML models, other features such as eccentricity and compactness were comparable. This finding does offer interesting implications for our understanding of microfracture and pore shape. Overall, these findings indicate that the microfractures in the dataset were sufficiently complex that a shape-based, object-classification approach within a supervised ML context offers a more potent solution than simply using aspect ratio. We believe that the proposed approach can potentially be helpful in several fields facing this challenge, such as mining, hydrogeology, and geotechnical engineering.

Dismiss