Class-Based Machine Learning for Intelligent Reservoir Characterization Over the Life Cycle of a Field in the North Sea

A consistent approach to evaluate the mineralogy and petrophysical attributes throughout the extent and life of the field is essential. This ensures proper assessment of the reservoir potential and provides deeper insights about the sedimentological and depositional environment. However, different logging technologies with varying capabilities and applications are often used during field development phases, introducing differences and discrepancy in the assessment. In a study from North Sea, only few wells had high-end spectroscopy measurements to decipher complex mineralogy. Due to the lack of similar measurements in other wells, we explore the application of Intelligent Reservoir Characterization using Class-based Machine Learning (CbML) to provide a reliable and consistent evaluation. During training phase of CbML, key wells with complex mineralogy evaluated using advanced wireline spectroscopy logs are used. First training data is automatically reduced into interpretable facies using a novel Petrophysical Data-Driven Classification (PDDC) methodology. Then facie-wise learning models, including 1)outlier detection, 2)prediction of target complex mineralogy from input spectroscopy, 3) feedback QC loop of reconstructing spectroscopy from learnt mineralogy, and 4)corresponding uncertainties, are developed. During prediction phase, learnt models are applied to basic spectroscopy data in non-key wells to automatically identify outliers, assign facies, predict target complex mineralogy with uncertainties and reconstruct spectroscopy for a comprehensive QC. The procedure takes advantage of traditional petrophysical workflows as well as machine learning algorithms to quickly assess and deploy an intelligent reservoir characterization application. The methodology was used over a blind well where the predicted mineralogy showed a good match with that of an expert driven assessment using high-end spectroscopy data. In the non-key well, the mineralogy predicted from the CbML application was used to determine different petrophysical attributes like porosity, intrinsic permeability, and elemental concentrations, which matched well with the conventional core analysis and XRF data. Overall, the uncertainties estimated from the workflow were very low, establishing further the robustness and reliability of the results. While it is desired to have all the necessary measurements to address the formation complexity, the well logging program is mostly constrained by the logging environment or planned borehole trajectory. The case study above shows how the CbML workflow can be used to develop a fit-for-purpose intelligent solution as an alternative to traditional interpretation. Once a learnt model is developed, it can be shared among experts and applied to any new well to provide near-instant, consistent quantities of interest, such as mineralogy, grain density and elemental concentrations as in our present case.