New analytical techniques are transforming the way that reservoirs are discovered as well as characterized and produced. With all the new processes and approaches, data continues to be a challenge, especially in developing reservoir models that are accurate and can be scaled.
Welcome to an interview with Benmadi Milad, Ph.D., who will be participating in the AAPG Success with Difficult Times GTW in Houston, Nov 12-13. His focus is on machine learning for better modeling using a wide array of data.
What is your name? What is your background?
I would like to thank you for the opportunity to participate in this interview! My name is Benmadi Milad. I received my Ph.D. (2019) in Petroleum Geology and M.Sc. (2013) in Petroleum Engineering from the University of Oklahoma (OU). Prior to that (2008-2010), I worked for Eni Oil and Gas Company in an offshore oil field. In addition to these experiences, I am a recipient of the First-Place Ph.D. Researcher, the Second-Place Ph.D. Researcher by ConocoPhillips Company, the Best Applied Geology and Geophysics by School of Geosciences at OU, and finally a recipient of the Distinguish Graduate M.S. Student by School of Petroleum Engineering at OU.
Throughout my academic career, I mentored undergraduate students for their B.Sc. honors thesis and worked on many projects as a team at the Institute of Reservoir Characterization (IRC), directed by Dr. Roger Slatt. While I was at the IRC, I led core workshops and field trips for graduate students and 28 oil companies, which are members of the Woodford Shale Consortium.
My research has been particularly focused on exploration and development of conventional and unconventional oil and gas reservoirs, using multi-scalar scientific data integration and using various statistical analysis and machine learning techniques. The emphasis of my research is on the integration of the interdisciplinary between the geology, geophysics and engineering data (i.e., wireline logs, outcrops, core data, seismic attributes, natural fracture, hydraulic fracture, and production data) to characterize and develop static and dynamic reservoir models for successful oil and gas exploration and development programs.
What do you consider to be a key challenge in today's oil and gas exploration and operations?
As science progresses and data collection increases, our ability as humans to interpret and classify that data is limited because of the immense amounts of information. Hence, the next step is Machine learning (ML) and Artificial Intelligence (AI). ML and AI stand primed to offload a lot of the work and construction of models automatically is possible. However, the accuracy of these models might be questionable if scientists and engineers do not consider their validation against rock (core and/or outcrops) and production data. The challenge is how we can assure that the subsurface lithofacies model fits actual rocks and their properties on log signatures and production performance. In addition, when we deal with outcrop data and sedimentological information, and being able to quantify that, it feeds directly into workflows for better classifications in the subsurface. An effective facies classification is critical for the reservoir characterization to reduce the risk/inaccuracy of facies redundancy and misinterpretation for an exploration and development programs, which could lose millions of dollars when you target inaccurate lithology especially when the stratigraphic thickness of the target zone is small.
Besides that, there are certain problems within reservoir modeling that need to be addressed, such as: limitations with grids, the ability to capture the heterogeneity observed in outcrops, and the level and scale of heterogeneity that matters for fluid flow, and the complexity of the structural and stratigraphy geology. For example, the minimum limit of heterogeneity that matters to fluid flow with the optimal grid sizes and grid types. There are future research directions in striving to solve some of those problems.
The challenge today is to upscale the static geological reservoir models without sacrificing the original structures or eliminating the rock properties and the layered nature of the reservoir. This must be carried out without omitting the important effects of reservoir heterogeneity, and without losing the geological features of the geological model. Losing these features could impact reservoir performance significantly and give false indications of fluid flow performance in the upscaled model. Therefore, one of the research questions to be answered is what is an optimum upscaling that can be done to depict the complexity and the geology of a reservoir?
Which technologies are addressing the issues?
Data integration of surface and subsurface helps build better predictive subsurface models. It is very useful to convert the outcrop studies into numerical data for subsurface study to evaluate geological assessments through the depositional environment and lithology. Using a uniform statistical methodology of classifying facies resulted in consistency within the outputs of optimal number of clustering and facies (Chemo and electro facies) outputs.
Have you done anything personally to adopt or develop technologies? What did you do? What was your goal?
What were the results?
For machine learning, identifying the number of facies/clusters from well logs and/or x-ray fluorescence (XRF) data are required prior to applying Self organizing map (SOM) and K-means clustering techniques. An arbitrary number of clustering can lead to redundant number of clustering of facies. Mechanisms for clustering decision making can be either visually by interpreter or numerically by the computer (supervised vs unsupervised). Facies classifications by interpreters is dependent on the person either splitter or lubber type. This might result in either too many or too few clusters and lead to misclassification of facies. Therefore, as previously mentioned, using a uniform statistical methodology of classifying facies results in a consistency within the output’s classification outputs.
A machine learning workflow that I’ve developed and successfully applied on Hunton Carbonate, McAlister Woodford Quarry, and I-35 Mississippian Sycamore Outcrop. In the Mississippian case, the I-35 outcrop Sycamore lithofacies can be quantitively predicted on any wireline log suite. Thereby, the outcrop findings are used to develop criteria for predicting rock properties from wireline logs since most companies acquire conventional logs. This ultimately helps predict all Sycamore rock types for exploration and/or development programs.
Anytime geological data is being quantified and started feeding into machine learning workflows and taking into account the knowledge we gain from cores, outcrops, logs, and seismic data into the subsurface, subsurface reservoir can be better predicted.
For the reservoir modeling issues, I integrated data at different scales from cores, logs, outcrops to seismic volume to develop a high-resolution stratigraphic framework and complex structures to build detailed fine grid cell size static geological models consisting of lithofacies, natural fractures, porosities, permeabilities, and saturation. The resulting detailed (20x30x1ft grid cell size) reservoir model was upscaled to 100x150x1 ft grid cell size for flow (reservoir) simulation. Also, I used a dual porosity/dual permeability model for higher accuracy even though the calculations require higher time for upscaling and flow simulation.
I’ve developed practical aspects of upscaling geocellular geological models for reservoir fluid flow simulations to find the optimum level of upscaling. The suitable level of upscaling is determined based on the comparison of history matching, the desired numerical accuracy, and the computational time. It is concluded that vertical upscaling is preferred to horizontal upscaling, as it preserves the numerical accuracy. Also found that natural fracture length has a significantly larger impact on production profile than the fracture width. Moreover, the study demonstrates that both the uncertainty in reservoir description and the error associate with the numerical simulation/upscaling affect the quality of prediction. Therefore, the source of error should be well understood.
What do you have planned for the future?
In combination with a solid geology and engineering background, I plan to work on interdisciplinary gaps. I set, fill, and role in the gaps between geoscience and engineering disciplines is one of my unique.