English abstract
During the exploration phase, understanding the reservoir is a demanding task due
to geological complexities and limited data availability. Advanced techniques are
used by geoscientists to interpret available data, such as well logs and seismic data,
for gaining insights about porosity, permeability, lithology, and hydrocarbon
saturation of the reservoir. Seismic inversion is one of these key techniques that
coverts seismic data into elastic properties like P-impedance and S-impedance.
This technique performs excellently given the seismic data is high-quality and
minimally corrupted by noise and multiples. However, its efficiency deteriorates
with poor quality seismic data. Machine Learning serves as a viable alternative in
these situations, offering the potential to transform seismic data into elastic and
physical properties, showing satisfactory results and performance in many
instances.
The Gharif Formation is a significant oil and gas reservoir producer in Oman,
primarily comprising sand and shale with a carbonate layer easily identifiable in
seismic data. However, differentiating between sand and shale layers within the
Gharif Formation proves difficult due to its depth and low seismic resolution
within its range. Furthermore, distinct differences in physical and elastic
properties make it especially challenging to visualize channel extensions and
segregate sand from shale in the Gharif Formation.
This project aims to leverage Machine Learning and Deep Learning capabilities
to convert seismic data into elastic properties, P-impedance, S-impedance, and
Vp/Vs, facilitating the differentiation of sand from shale in the Gharif Formation.
Several algorithms were tested to achieve optimal results, including eight key
algorithms: shallow algorithms like Linear Regression, Ridge Regression, Lasso
Regression, Support Vector Machine Regressor, Rain Forest Regressor and Multilayer Perceptron, and Deep algorithms such as Convolutional Neural Networks
and Deep Neural Networks.
The study was divided into three stages. The first stage involved using various
algorithms to transform synthetic seismic data generated from wells into Pimpedance. The shallow algorithms' results weren't as effective as expected,
achieving a correlation of about 70%. Conversely, the deep algorithms
demonstrated excellent results, with DNN, CNN, and the shallow algorithm MLP
achieving 89.65%, 89.09%, and 76.05% of accuracy respectively. In the second
stage, MLP, DNN, and CNN algorithms were compared regarding their
performance in predicting P-impedance from actual 3D full-stack seismic data,
with MLP outperforming the rest. The project's final stage involved using substacks of seismic data (near, near-mid, mid and far) to predict P-impedance, Simpedance, and velocity ratio. The AI-based inversion results were validated using
seven blind wells, showing a similar performance to the science-based AVO
inversion. Subsequently, the AI-based inversion outputs were used to calculate
sand and shale probabilities using the rock fluid index (RFI). The probability cube
enabled the extraction of several maps below the upper Gharif horizon to highlight
the channels.
