Welcome to RokBites, your bi-weekly guide to brushing up on RokDoc techniques. 👩💻💡
This week we’re diving into machine learning and its use in incorporating various rock physics models for reservoir property prediction.
Integrating Rock Physics with Machine Learning (RPML)
Deciding a Rock Physics (RP) model in a given depositional setting can be a daunting task when you have limited to no core information available. RokDoc is at the forefront of innovation by integrating machine learning into its workflows. The Rock Physics Machine Learning (RPML) toolkit uses the expectation-maximisation algorithm in a data and RP driven context, to predict facies and petrophysical logs in multi-well scenario. This toolkit provides a head start in creating Rock Physics model templates which can then be shared easily among teams and represent the base start for further refined calibration.
Application in US Onshore Basin
Midland Basin, a sub basin of Permian, is known for its rich oil and gas reserves, once considered unrecoverable due to tight rock formations. Channel deposits, mass transport complexes, tectonic events lead to varying degrees of sediment deposition and mineralogical variations. The basin stratigraphy is widely characterized as a product of complex depositional processes with varying degrees of detrital limestones, siltstone, mudstones, calcareous and organic rich shales having multi-benched stacked pays. Depending on the location within the basin, some of these facies can be drilling hazards while others can act as a primary source for hydrocarbon generation. Therefore, accurately predicting and mapping facies distribution is key to successful reservoir characterization workflow.
For our study, we look at a few key sections from from the Devonian to Grayburg units:
And this case, there are some petrophysical data to help validate the RPML prediction.
The units of particular interest, which were evaluated by Wilson et. al (2020) with respect to the depositional environment of the lower Spraberry and Wolfcamp B, cause a challenge in that there is high vertical and lateral variability in the mixed carbonate and fine-grained clastic sequences.
The muddy siltstone, Facies 2 in the geological map, (avg TOC = 2.5 wt%) appears to be the most optimal reservoir with high TOC, low clay content, and high quartz + feldspar content. Facies 1 - "silty mudstones" (avg TOC = 3.6 wt%) and Facies 3 - Silty calcareous mudstone (avg TOC = 2.9 wt%) also constitute good/fair reservoir potential.
https://doi.org/10.1306/12031917358
RPML Calibration
Designing an RPM for different zones and multiple facies can be time consuming. For example, within the Midland Basin, 5 stacked pays with 3-4 facies classification means calibrating 15-20 RPMs within each zone. RPML takes that pain point away and incorporate petrophysical and pressure data to optimize for facies and elastic trends using various rock physics models and optimizes their parameters simultaneously.
The inputs include petrophysical and elastic logs, which can be left blank if they are not available. RPML will use data where available to predict missing information for each well.
The user can pick an RPM from a library of models available. In this example, the template characterizes facies, petrophysical and elastic trends along with vertical effective stress (VES) using the Vernik-Kachanov Sandstone, Shale and Carbonate RPMs. The following rock types are characterized along with their prior probabilities:
Rock Types: Dolomites, Siltstones, Organic Rich Shales, Calcareous Shale, Carbonates
Interval of Interest:
- Grayburg: Contains 27% Siltstones, 60% Carbonates, and 3% Grayburg Dolomites
- San Andres: Contains 68% Siltstones, 26% Organic Rich Shale (Shallow), and 5% Carbonates
- Leonardian: Contains 85% Siltstones, 10% Organic Rich Shale (Shallow), and 5% Carbonates
- Wolfcamp: Contains equal portions of Organic Rich Shale (Shallow), Calcareous Shale, and Carbonates
- Barnett: Contains equal portions of Organic Rich Shale (Deeper), Calcareous Shale, and Carbonates
Predictions (red) vs measured log data (black) for one of the wells along with Facies classification.
Blind well results:
The RPML algorithm is applied to a blind well to test the robustness of the model. In this case Vs, Rho and GR are predicted from the training.
The toolkit also gives per-facies depth trend, calibrated RPMs and prior-facies distribution which can be utilized in other reservoir characterization workflows, such as Ji-Fi, Depth-varying Bayesian Classification, and AVO modeling.
Conclusion
RPML is a powerful tool that ‘intelligently automates’ facies clusters based on the log response. Depth trends are outputted, as well as calibrated RPMs – kickstarting the interpretation workflow. The outputs can be used to understand AVO effects, patch missing log information and use the predicted results in inversion techniques.
Give it a try!
This study has been encapsulated as an RPML template in RokDoc. Simply reference your project working intervals to those found in the template description, mineral and fluid sets will be automatically imported, and the parameter ranges for each rock physics model will be optimized with your project data!
We hope you found this post insightful. Feel free share your feedback and propose any topics you would like us to explore in future posts. Your input helps us create content that truly resonates with our community.
Thank you for being part of our journey - see you next post! ☕🍪
