Supervised machine learning sedimentological characterization workflow: A tool to bridge the gap between qualitative and quantitative facies interpretation and uncertainty quantification
Anis Seksaf, Boris Kostic, Chloé C.F. Château, Daniel Clay, Kamel Tamene, Meriem Bertouche, Jan Van Der Wal, Raja Ramalingam, Abd El-Aziz Sabry, Boland Ghadeer Taleb, Chao Chen · 2024
From pioneer exploration geologists to modern upstream reservoir characterisation asset teams, the geological interpretation of petroleum potential in uncored wells has always been a challenge, despite the enormous technological advancements. Many attempts have been performed through the years to try and bridge the gap between cored and uncored domains, and more recently Artificial Intelligence (AI) has proven to be an effective addition to the conventional workflows. Machine learning algorithms were built to mimic the way that the human brain works. The basic idea behind an AI algorithm is to: i) give it enough information so that it learns the way humans do, without being explicitly programmed ii) improve its learning over time, in an automated way. It learns from patterns, historical records and events, similar to the way that humans learn. In this reservoir characterisation project, the dataset that is used to train the algorithm is the sedimentological interpretation of the geologist. Humans make decisions based on past experiences, such as feeling pain when touching a hot surface, reading about lions and knowing they can attack humans, and hearing a loud noise and seeing a car crash. In geoscience, the learning process is based on acquired theoretical background knowledge and hands-on experience from geoscientists. A facies identification, based on core, begins with sedimentological core description. For every facies described and interpreted by the sedimentologist, there are associated wireline and/or LWD tools that provide a quantitative measurement. A computer can be programmed to implement algorithms that output predictions in a quantitative way, providing information that could be used to help solve problems, such as prediction of geological parameter in uncored wells. Before machine learning, traditional computer programming performed actions based on conditional statements such as “If”, “Then”, “Else” rules, and this could be identified as ‘cut-offs’ for wireline logs, where some facies have clear log responses. Even so, geology is very complex and often the basic cut-off relationship shows limitation, especially when several inputs are to be considered. Algorithms are the core of machine learning models and are based on sets of rules designed from the input/output implicit relationship. A machine learns when historical data are provided as input, such as past interpretations. It can then look at patterns in the data to learn rules. These are the foundations of an AI based algorithms. The study aims at showcasing the application of AI to the sedimentological and stratigraphic characterisation of a Lower Cretaceous formations in a giant oil field in the Middle East. This is primarily based on the detailed depositional evaluation of cored wells and borehole image logs, supplemented by previously described historical cores. A consistent descriptive and stratigraphic framework has been applied across this dataset and extrapolated into uncored intervals and wells utilising supervised machine learning to predict facies distribution. This document specifically examines the details of machine learning aspects and algorithms’ construction, as well as workflow implementation and value creation resulting from integrating a quantitative dataset with an initial fully qualitative interpretation. The dataset is based on 14 wells (covering different formations), 9 cored and 7 uncored but with borehole image data available. Only 2 wells have both core and image data. Furthermore, 12 historical cores, together with 148 uncored wells, were also included in the study (prediction dataset). The consistent descriptive and interpretative sedimentological scheme applied across the cored and imaged wells was extrapolated into the uncored intervals and wells using the validated supervised machine-learning model. The studied interval comprises three main formations: Zubair (Azim et al., 2019), Ratawi Shale and Ratawi Limestone. Due to the geological differences among the studied formations, a specific supervised machine learning model has been developed for each of them. The construction of the model has been strongly and strictly dictated by geological rationale throughout the process. The prediction from the model provided quantitative data points, which helped to enhance and improve the overall mapping of geobodies, while allowing a quantification of the geological uncertainty.