3D and 4D inversion for rock and fluid properties using deep learning

Abstract

This thesis focuses on estimating rock and fluid properties from the perspective of 3D and 4D seismic inversion. I developed two techniques that enable seamless integration of 3D and 4D seismic data. The first technique emphasises the estimation of porosity, Vclay, and hydrocarbon saturation directly from 3D seismic data using deep learning. Additionally, I propose an approach to enhance the lateral continuity of these estimated petrophysical properties. The products from this first technique are subsequently integrated into the 4D domain, leading to the development of the second technique that enables the inversion for reservoir pressure and saturation changes from 4D seismic data using deep learning. Both techniques involve the use of synthetic training datasets for network training, where the detailed processes for building realistic training datasets are presented. The first technique was tested across four fields with diverse deposition environments, covering meandering fluvial systems, fluvial estuaries, deepwater settings, and carbonate platforms. The second technique was applied to the meandering fluvial field with available 4D seismic data. This technique successfully distinguishes pressure effects from saturation-related effects in the 4D seismic response. It also highlights the importance of incorporating fluid flow information into the training dataset, enabling the network to capture the relationship between the superimposed effects of dynamic property changes and the corresponding 4D seismic response. Finally, I present a summary of the cost-benefit analysis of these developed techniques, demonstrating their ability to accelerate the inversion process in terms of turnaround time while providing robust solutions when applied to field applications.