Rock Physics Driven Machine Learning for Quick & Improved Reservoir Characterization

Machine learning has been used in the petroleum industry for a long time, but its usage was limited due to hardware and data constraints. With the advancement in hardware capabilities in recent times, machine learning usage has expanded in various domains. Still, in many real situations, the inadequacy of well data required in seismic reservoir characterization poses a challenge to the use of recently developed deep machine learning methods, e.g., convolutional neural networks (CNN). Theory guided machine learning (TGML) generates a large amount of 1D synthetic data to capture the variability in the conditions of the reservoir using a rock physics model, conforming to the regional geology and depositional setup. The corresponding amplitude variation with offset (AVO) responses are used for training and validating a CNN network. The concept of transfer learning is used to validate the CNN training on real well properties before applying to the 3D seismic data for predicting several elastic and reservoir properties simultaneously. Here, we present a case study on a West Tryal dataset from the Northern Carnarvon basin, Australia with limited well control in the survey area. A rock physics model is established on one of the wells and then geological knowledge about the area is used to simulate various scenarios of reservoir variation in subsurface to predict the elastic properties in 1D. Each set of reservoir and elastic properties can be regarded as a synthetic well. A real-world wavelet is used to compute the AVO responses for each synthetic well. With this, there are many synthetic wells and synthetic seismic data to be used in the deep neural network for machine learning. Trained and validated convolutional neural network is then transferred on the real dataset and later applied on the 3D seismic data to predict multiple reservoir properties, acoustic impedance, Vp/Vs, porosity, volume of clay and water saturation simultaneously. A comparison is made between the acoustic and density prediction from seismic inversion, and one predicted from TGML and between porosity and water saturation predicted from a conventional workflow and the one predicted from TGML, showing the improvement in quality of prediction and value addition by removing workflow repetition.