Reservoir Property Distribution Constrained with 3D Seismic Data in a Tight Carbonate Reservoir: A Machine Learning Based Approach

Abstract

When dealing with a tight reservoir of basinal settings, depositional facies can show a lot of overlap in their properties. This onshore tight carbonate reservoir, is almost exclusively composed of wackestones to mudstones. The main reservoir quality influencer is the diagenetic overprint reflected in the cementation and dissolution phases, and directly influencing the porosity variation. This made the classical generation of facies trend maps to constrain the properties less reliable, due to poor property discrimination. It demanded the utilization of seismic data to attempt to constrain the properties as much as possible. With the static rock types being sensitive to porosity variation, and porosity having a good correlation to the acoustic impedance, the seismic data may provide key information to obtain lateral trends guiding the properties distribution. A supervised neural network approach was used to provide a link between the lateral variation of reservoir quality and seismic data. The workflow relied on combining deterministic inversion and volume attributes to extract the information. This was followed by multi-attribute analysis to find the best combination of attributes to input into the neural network. Four main challenges had to be tackled. Firstly, an acquisition footprint masking the signal in the seismic data had to be reduced. Secondly, the integration of the seismic inversion output and well information in the modeling workflow. Thirdly, generating and finding the best combination of seismic attributes to fit the given logs. Lastly, adapting the low resolution of the seismic data to the subzones of the reservoir in order to capture the vertical variation. The resulting predicted impedance cube is validated with actual porosity data from wells. The result has proven to be beneficial in building an integrated 3D property model, giving more confidence in the predictability of the properties in horizontal wells and volume estimates. This is seen in the blind tests applied, which show a good alignment between actual and predicted properties.