Machine Learning Techniques Applied to Waveform Inversion for Porosity Estimation in Hydrocarbon Reservoirs

• Grant number: 23/17610-8
• Support Opportunities: Scholarships in Brazil - Doctorate
• Start date: April 01, 2024
• End date: January 31, 2028
• Field of knowledge: Physical Sciences and Mathematics - Geosciences - Geophysics
• Principal Investigator: Roberto Hirata Junior
• Grantee: Janaína Anjos Melo
• Host Institution: Instituto de Astronomia, Geofísica e Ciências Atmosféricas (IAG). Universidade de São Paulo (USP). São Paulo , SP, Brazil
• Associated research grant: 22/15304-4 - Learning context rich representations for computer vision, AP.TEM

Abstract

The characterization of reservoirs is an essential stage in the development, monitoring, and management of a hydrocarbon reservoir, as well as in the optimization of the production of important components for human use. An important physical parameter of the reservoir is porosity, as it reflects the quantity of oil and gas present in a reservoir. Thus, accurate calculations of reservoir porosity are crucial in geological interpretation, exploration, and reservoir development. Porosity can be estimated through seismic tomography, a data inversion technique that allows imaging the velocity or attenuation structure of the Earth's interior. With the increasing computational power over the last decade, numerical techniques solving the equation of motion for elastic and poroelastic media can now be applied to the study of media with complex physical properties and structures. Owing to the need for increasingly accurate and higher-resolution elastic property models, the Full Waveform Inversion (FWI) method, another way of performing tomography, has been increasingly applied. FWI is a nonlinear inversion process that minimizes the difference between the waveform recorded at a receiver and the waveform calculated by a physical model to estimate elastic parameters of a medium. It allows exploring all finite-frequency effects contained in a seismic record and provides proven models of physical parameters with higher resolution. Thus, elastic parameter models provided by FWI have been increasingly used for estimating the physical properties of rocks for reservoir characterization. Machine Learning techniques have several advantages over traditional interpretation techniques, as they enable connecting usually unrelated parameters and solving common nonlinear problems when trying to convert geophysical data information into geological information or reservoir petrophysical properties, with more accuracy and less interpretative bias. Given the ability of Machine Learning to handle nonlinear data without the need to specify the precise nature of nonlinearity a priori, this project proposes the use of two Neural Networks (supervised and unsupervised) for the processing of seismic signals and interpretation of images generated by FWI. This procedure aims to estimate relative porosity at multiscale, integrating three different study scales (seismic tomography, well logs, and laboratory core measurements) to improve the accuracy of the spatial distribution of this petrophysical property. To achieve this, synthetic and real data will be used to train the Neural Networks, adapting existing codes in the Python programming language and developing new codes when necessary. Therefore, the Neural Networks resulting from this work will be trained to have the ability to generate interpretation and porosity estimates for a reservoir at three different study scales (seismic, well, and laboratory), developing the capacity to interconnect these different scales of details when this information is available in the same study area.