Data science-based methods, such as supervised neural networks, provide powerful techniques to predict reservoir properties from seismic and well data without the aid of a theoretical model. In these supervised learning approaches, the seismic-to-rock property relationship is learned from the data. One of the major factors limiting the success of these methods is whether there exists enough labeled data, sampled over the expected geology, to train the neural network adequately. To overcome these issues, this paper explores hybrid theory-guided data science (TGDS)-based methods. In particular, we build a two-component model in which the outputs of the theory-based component are the inputs in the data science component. First, we simulate many pseudowells based on the well statistics in the project area. The reservoir properties, such as porosity, saturation, mineralogy, and thickness, are all varied to create a well-sampled data set. Elastic and synthetic seismic data are then generated using rock physics and seismic theory. The resulting collection of pseudowell logs and synthetic seismic data, called the synthetic catalog, is used to train the neural network. The derived operator is then applied to the real seismic data to predict reservoir properties throughout the seismic volume. This TGDS method is applied to a North Sea data set to characterize a Paleocene oil sand reservoir. The TGDS results better characterize the geology and well control, including a blind well, compared to a solely theory-based approach (deterministic inversion) and a data science-based approach (neural network using only the original wells). These results suggest that theory and data science can complement each other to improve reservoir characterization predictions.

You do not currently have access to this article.