Machine Learning Provides Higher-Quality Insights into Facies Heterogeneities over Complex Carbonate Reservoirs in a Recently Developed Abu Dhabi Oilfield, Middle East

Abstract

Because of the complexity of properties and heterogeneities, the challenge in a carbonate reservoir is to predict the spatial distribution of the best reservoir facies. Due to the sparse distribution of wells, uncertainties exist, especially where fewer cored wells are available. The aim of this study was to employ machine learning, using the full dimensionality of 3D seismic data and well data, to predict lithofacies heterogeneities distribution in major reservoirs of the Thamama Group, for a recently developed large UAE onshore field. This technology generates a probabilistic seismic facies model derived from the 3D seismic data. An association of naive neural networks, each with a different learning strategy, is run simultaneously, to avoid biasing any of the neural network architectures. To train the neural networks, seismic data and the lithofacies at the well location extracted along the wellbore are used as labelled data. To avoid overfitting from a limited dataset, we introduce seismic data away from the borehole (soft data) so that the neural networks can "vote" on their integration to improve the final training dataset before reaching the ultimate learning stage. The application of this technique on Lower Cretaceous carbonate reservoirs shows promising results. The analysis of the probability distribution gives good insights into reservoir facies distribution uncertainty. Lithofacies are created from electrofacies by subdividing facies based on hydrocarbons. The resultant prediction was validated through comparison with observations from a new drilled well, adding confidence in the decision-making process when selecting future drilling locations. This method uncovers new potential for seismic data reliability when predicting the reservoir lithofacies away from wells, especially when referring to prestack data with any type of seismic attributes. Using this method, the major reservoir lithofacies can be precisely predicted within the field. As the probabilistic facies model is calibrated to wells, this lithofacies data can be used for both geologic modeling and volumetrics analysis. Machine learning techniques were successfully applied to generate lithofacies from electrofacies from the 3D seismic data, leading to accelerated interpretation and reservoir characterization processes. In many cases, they provided faster images of the subsurface while still maintaining accuracy, thus helping to improve the decision-making process when determining new drilling locations.