TUESDAY, 24-JUN-25 12:53
TUESDAY, 24-JUN-25 12:53
"Machine Learning for Rock Typing"
Traditional rock typing was performed either indirectly via petrophysical properties measured from downhole logging tools or directly by physical examination and analysis of core samples. Knowing the rock type allows a more accurate estimation of rock and/or fluid properties within a formation which in turn enables a better prediction of its potential value as a mineral deposit or hydrocarbon reservoir.
Technological innovations in both well logging tools and core logging and imaging solutions now deliver much more data than was available traditionally. This dense, high-resolution data can be analysed using advanced machine learning (ML) techniques that translate patterns within the data into rock types far more efficiently than could a human interpreter.
This webinar by the Rock Imaging Special Interest Group (RISIG), the organisers of the highly regarded International Rock Imaging Summit, will bring together a group of experts in the field who will present firsthand accounts of the challenges and benefits of implementing reliable ML implementations. There will be opportunities for attendees to question the speakers individually and as a group.
REGISTRATION IS CLOSED.
Members who register for the webinar will be able to view the recorded presentations after the event by logging in and clicking on the title of the talk opposite.
Agenda - Wednesday 8th September 2021 - all times are GMT.
Talks will be 20 minutes long followed by 10 minutes of Q&A.
BIO: Alfarisi is an Adjunct Professor on Digital Rock Typing and Reservoir Characterization since 2019, and Research Project Champion of "Machine Learning for Fluid Flow" collaborative project with Stanford University at the Department of R&D of ADNOC. Alfarisi is the 2020 Keynote Speaker of the National Competition of Robotics & Artificial Intelligence, UAE.
Alfarisi is a Reviewer at SPE (Society of Petroleum Engineers) Formation Evaluation Journal. Alfarisi is co-Chair of Reservoir Characterization and Modeling Conference themed Digital Transformation. Alfarisi is an SPWLA (Society of Petrophysicists and Well Logging Analysts) Distinguished Speaker on Carbonate Rock Typing in 2017 and 2018, experienced with Fortune Global 500 companies, including BP, TOTAL, and Schlumberger. Alfarisi is six times the winner among 100's innovation of the ADNOC Offshore Innovation Award.
Alfarisi received the 2016 "Most Likely to Change the World Award" from London Business School Student Association. He taught courses and mentored graduate and undergrad of ~148 students in Asia and the middle east, China University of Petroleum, and Khalifa University of Science and Technology. His research passion is Deep Learning for 3D Digital Rock Typing.
SUMMARY: Rock physical and chemical properties are the main features that determine the rock type. Lithology, permeability, pore throat size, capillary pressure curve are the properties that classify the carbonate rock. To perform properties identification manually is time consuming and can be very length process that all geoscientists suffer from. We built a rock typing framework that utilizes the latest computer vision advances in analyzing digital rock through semantic segmentation of micro-Computerized Tomography (uCT) images and the Carbonate Morphology Chart (CAMO-Chart), and we called it Digital Rock Typing (DRT). We provide geoscientists with DRT to be efficient, consistent, and have a higher-quality rock classes.
BIO: After studying structural geology and geophysics at Monash University, Brenton has consulted as a structural geologist and geophysicist for PGN Geoscience as well as working in a variety of geological and geophysical roles—predominantly in exploration. Brenton has also worked as a geophysicist and data scientist for MMG Exploration working in Nickel, Copper and Zinc exploration and project generation.
In 2015, Brenton co-founded Solve Geosolutions - an exploration and mining focused data science consultancy where he currently serves as Director. In 2019, Brenton co-founded Datarock - a computer vision technology company geared at building productionised image and video analysis solutions for exploration and mining.
SUMMARY: With recent improvements in deep learning based image analysis, we now have the technology to build models that can automate visual tasks previously deemed to be too difficult.
Datarock has built an image analysis platform that can generate complex geotechnical datasets such as Rock Quality Designation (RQD) and the detailed analysis of joints and fractures automatically from both traditional core photography and 3D scanning datasets such as LIDAR.
In this talk we demonstrate the technology and workflows created by Datarock, as well as showing how they are being used in the minerals industry to improve orebody knowledge.
BIO: George has 8 years of E&P experience from geoscience roles in exploration, development, and research projects, working in Malaysia, Austria, and Norway. Trained as a geologist with an open mind for geophysical applications, he has been specialising in the quantitative interpretation of carbonate reservoirs, with a particular interest in litho-facies inversion and classification algorithms. George holds Master’s degrees in Applied Geoscience and Business Administration from the Universities of Leoben and Munich, and has a proven track record of leading cross-functional teams, results driven, in challenging projects.
SUMMARY: Predicting lithology from well logs in a complex geological setting is a non-unique classification problem. In this Western Barents Sea case study, we aim to reduce uncertainty by applying an optimised labeling strategy to Machine Learning (ML) assisted lithology prediction in order to obtain geologically plausible results. The region is characterized by the occurrence of two upper Palaeozoic carbonate factories and a range of associated breccias and conglomerates that document considerable tectonic activity. Multiple rifting events created horst and graben morphologies that facilitated carbonate platform development along structural highs. Subsequent basinward tilting and repeated subaerial exposure led to karst and collapse features and the potential to develop mass flow complexes with a broad range of composite lithologies in redeposited sediments.
In this study, we focus on core samples from the upper Palaeozoic Ørn formation, which is characterized by bioherm - style carbonate build-ups with algae and crinoid rich patch-reef facies, as well as a range of collapse breccias and conglomerates resulting from tectonic activity, erosion, and subaerial exposure. For classification, ML models have been trained with wireline logs as features including density, resistivity, compressional and shear sonic slowness (DT and DTS), as well as gamma-ray (GR). We benchmarked algorithmic performance from Gradient Boost, Random Forest, and a fully connected neural net against a linear Ridge classifier as a baseline predictor. Results indicate that the neural net performs best in predicting small scale heterogeneities, with an average blind score of 60%, and a best individual class score of 89%, which has been achieved on brecciated rocks.
BIO: Gwenolé TALLEC has had his master degree from the University of Bordeaux, France in 2002, in computer science, image processing and 3D graphics programming. He focused in Digital Rock Analysis in 2010, and is one of the PerGeos Expert at ThermoFisher, delivering training courses, defining workflows at customer sites, and linking users and software’s R&D.
SUMMARY: With the emergence of digital rock analysis, its application to whole cores and derived samples is bringing a tremendous help in their analysis and characterization. While visualization of core data is indeed an appreciated support, derived digital properties bring crucial information to validate and enhance laboratory analysis. In this presentation we will go further. By applying feature augmentation Machine Learning models applied to whole core for facies identification. The main goal is to help petrophysicists and geologists to determine geological and petrophysical facies, by both bringing added value and filling missing information in their estimation, and applying a model defined on a subset to a whole well, in order to dramatically accelerate its characterization.
In this case study, we will apply the methodology to 35 metres of cored well data acquired from the Late Jurassic Upper Jubayla Formation, equivalent to the lower Arab-D reservoir in Saudi Arabia. The main goal of the study will be to bring Machine Learning models from original and derived digital rock properties, to automatically extract facies in the whole well, from a preliminary definition of the geological and petrophysical facies. We will specifically focus in this talk about pure Machine Learning results, and compare the final obtained scores depending on the type and number of variables we input. To conclude, we will expose and prove how Machine Learning models developed with Python using third party libraries, applied to the whole well, can automatically determine rock facies by augmenting the preliminary analysis by the integration of digital rock derived properties.
BIO: Laurent Bernard is the CEO of Reactiv'IP. Laurent Bernard founded the company Reactiv'IP in 2013 with the aim of offering new modern and fast tools for scientific image analysis.
SUMMARY: Recent available imaging technologies for rock and material analysis allow visualization of structures with great precision. But besides visualizing them, new Machine Learning tools enable a fast segmentation of most datasets acquired through micro-tomography, FIB-SEM microscopy, etc.
This presentation will present IPSDK Smart Segmentation modules suite through several practical projects performed by IPSDK users. The first example will present a segmentation of different types of materials present in an FIB-SEM image, from registration to STL surface file generation. The second example will show how Machine Learning tools can easily detect cracks and identify different materials in a rock slice.