Carbonates

Carbonate Reservoir Case Study

In our first carbonate reservoir case study, we tested whether our 3D Machine Learning Inversion (3DMLI) workflow based on ensemble machine learning could be implemented to characterize highly fractured, low porosity carbonates of Cretaceous and Jurassic age. The main reservoir intervals correspond to Cretaceous and Kimmeridgian carbonates that show favorable geological conditions to migration and the accumulation of hydrocarbons generated from the rich total organic carbon (TOC) Tithonian carbonates. These conditions are associated with the presence of fractures distributed in a heterogeneous manner within partial and sporadically dolomitized carbonates, especially between the upper and middle cretaceous.

A second interval of interest exists and is known as an Oolitic bank within the Kimmeridgian. It is constituted by dolomitized carbonates with relatively high effective porosity, predominantly related to inter crystalline and dissolution pore space development, sporadically affected by fractures.

The input data available to carry out the carbonate reservoir characterization includes prestack seismic angle stacks and two wells. The main logs for both wells are displayed in Figures 1 and 2.

Figure 1: Well A logs.

Figure 2: Well B logs.

Figure 3 shows a seismic amplitude spectrum comparison between the entire fullstack cube and a subset of the seismic focused on the zone of interest. Severe loss of frequency content is observed in the zone of interest (~5 – 17 Hz dominant frequency range) due to carbonate geology and depth below the surface. The data have a 4 ms sample rate, so theoretically, the seismic has up to 125 Hz of information.

Figure 3: Seismic spectrum comparison between the entire fullstack cube and the zone of interest.

The High-Definition Rock Property workflow carried out and details regarding the ensemble machine learning technology are discussed on the Machine Learning page. Figure 4 displays the High-Definition random line results and their respective well logs. Our prediction includes compressional velocity (Vp), density (RHOB) and gamma ray. A fullstack seismic random line is also shown for comparison. A high correlation between the measured and predicted logs is observed for both wells. Moreover, the work carried out has resolves fine-scale subsurface features that are not apparent in the seismic data alone.

Figure 4: Random line between wells A and B for fullstack seismic (Top Left), High-Definition density (Top Right), High-Definition compressional velocity (Bottom Left), and High-Definition gamma ray (Bottom Right).

In order to gauge how trustworthy the High-Def results are, blind well trace analyses are carried out (Figure 5). In the first analysis, Well A was used to build the machine learning model and applied to Well B, which serves as a blind well. This means the result at Well B has not been influenced in any way by the measured data at Well B. The same analysis is done by extracting the machine learning model at Well B, and using Well A as a blind well. Both results show an acceptable correlation at the blind wells, giving a high level of confidence in the work carried out.

Figure 5: Quality control analysis for gamma ray using blind wells.

Figure 6 shows a random line comparison between our High-Definition acoustic impedance and an inverted acoustic impedance using conventional seismic inversion based on Aki and Richards reflectivity approximation. The results clearly show an impressive difference between resolution, and this observation is corroborated by the impedance spectrum comparison in Figure 7. The fullstack seismic amplitude spectrum is also displayed. Notice the deviation between impedance datasets at approximately 23 Hz, which matches to where the a significant drop in seismic amplitude is observed.

Figure 6: (Top) High-Def acoustic impedance random line. (Bottom) Conventional seismic inversion result for acoustic impedance.

Figure 7: Amplitude spectrum comparison focused on the zone of interest for High-Definition AI versus AI derived from conventional seismic inversion. The fullstack seismic amplitude spectrum is also displayed.

A trace comparison for both wells is shown in Figure 8. The High-Definition comparison has a 120 Hz filter applied, while the conventional inversion comparison logs are filtered with a max frequency of 50 Hz. The High-Def AI shows a much higher correlation with the measured logs as well as an approximately 5x the resolution of the conventional seismic inversion approach.