Improving Reservoir Property Prediction Using Synthetic Data Catalog and Deep Neural Network in Poseidon field, Australia

The ultimate goal of reservoir characterization is to predict the distribution of elastic properties, porosity, and fluids in the target area. For many years Machine Learning techniques have been used in geophysics for different applications, including reservoir property prediction.

In these supervised learning approaches, the relationship for predictions is derived from the data. One of the major limiting factors for these workflows is the lack of labelled data covering the expected geology, therefore, it is challenging to train the neural network sufficiently. To overcome this, the hybrid Theory-Guided Data Science-Based method was applied.

The aforementioned workflow is divided into two main steps: first, generate many pseudo wells based on the statistics of the real well data in the project area. The reservoir properties, such as porosity, thickness, water saturation and mineralogy, are varied to cover different geological situations. Elastic properties and synthetic seismic gathers are then generated using rock physics and seismic theory.

The resulting set of synthetic data is used to train the neural network. The operator, derived during neural network training, is then applied to the real seismic data to predict properties throughout the seismic volume.