ABSTRACT

Determining litho-facies distribution is vital in reservoir characterization. It generally involves extensive analysis of well data and seismic data inversion. However, well logs analysis is limited to the well location and logs availability, while inversion is a very complex process that usually requires pre-stack seismic data, which is not always available. Seismic attributes can be easily extracted from any available seismic data and are very powerful in highlighting amplitude anomalies related to the elastic properties corresponding to lithologies, facies, and fluid contents. This study proposed methods for predicting seismic attributes to litho-facies by applying machine learning algorithms.

Seismic attributes are trained on four machine learning algorithms, namely, Random Forest (RF), Support Vector Machine (SVM), Gradient Boosting Tree (GTB), and Long Short-Term Memory (LSTM). Data used in this study consisted of seismic and well data from the Volve Field. Litho-facies interpretations are used as labels to train pseudo-wells of composite seismic traces. Additionally, AVO attributes are used for litho-facies prediction since partial stacks are available in the dataset.

The resulting models performed with generally high accuracies. It was found that the sampling rate of litho-facies logs, attributes selection, training intervals selection, and training sample size significantly affected the model result. Finally, the resulting models are used to predict the litho-facies distribution across several seismic slices, where they predicted geologically consistent sand distributions within the Volve’s field geological setting. For further improvements, more robust training datasets are needed for the model to increase prediction accuracy for each litho-facies.