Publications

Machine learning has been used for many decades, particularly for deep neural networks that were introduced in the 1960s and have been applied to various industry problems. In the past, machine learning techniques such as automatic amplitude picking, multi-linear regression and neural networks have been used routinely in seismic interpretation and reservoir property predictions. One of the main challenges of supervised learning in geoscience is the requirement of labeled data input for the training process. It becomes a major concern in areas that have a limited number of wells, as the existing data is not well sampled over the expected geological conditions. This paper explains the application of the hybrid theory-guided data science (TGDS) approach. Data augmentation is performed by simulating many pseudo-wells based on rock physics models (RPM) and well statistics in the project area. An appropriate RPM allows us to carry out systematic changes in reservoir properties (e.g., porosity, water saturation, volume of clay, and reservoir thickness) to create a well sampled training dataset representing the geological variations and heterogeneities. Elastic pseudo-well logs and synthetic seismic data are then generated using rock physics and seismic theory. The resulting collection of data, called the synthetic catalog, is used for training and validating the Convolutional Neural Network (CNN). CNN includes the process of transfer learning to validate the CNN training on real well properties before applying to 3D seismic data for predicting elastic and reservoir properties simultaneously. This TGDS method is applied to a West Tryal Rocks data set to characterize 3 main gas sand reservoirs in the Mungaroo formation. The TGDS indicates improved results even with a limited number of wells that characterize the reservoir properties and match to the well control, including a blind well. In addition, the TGDS results are compared to a theory-based approach (deterministic inversion). The results suggest that theory and data science can complement each other to improve reservoir characterization predictions.