Data Driven Analytics for Geological Storage of CO_{2}
Shahab D. Mohaghegh - May 2018
Data-driven analytics is enjoying unprecedented popularity among oil and gas professionals. Application of this technology to geological storage of CO_{2} is the focus of this book. Many reservoir engineering problems associated with geological storage of CO_{2} require the development of numerical reservoir simulation models. This book is the first to examine the contribution of Artificial Intelligence and Machine Learning and data-driven analytics to fluid flow in porous environments, including saline aquifers and depleted oil and gas reservoirs. Drawing from actual case studies, this book demonstrates how smart proxy models can be developed for complex numerical reservoir simulation models. Smart proxy incorporates pattern recognition capabilities of Artificial Intelligence and Machine Learning to build smart models that learn the intricacies of physical, mechanical and chemical interactions using the type of numerical simulation and modeling that is commonly used in the oil and gas industry. This groundbreaking technology makes it possible and practical to use high fidelity, complex numerical reservoir simulation models in the design, analysis, and optimization of carbon storage in geological formations projects. Following is a brief description of Smart Proxy Modeling that has been presented in this book.
Smart Proxy Modeling
There are well-known facts regarding the shortcomings of the numerical reservoir simulation and its practical applications that need to be addressed if this tool is to be effectively used for the geological sequestration of CO_{2}. These shortcomings can seriously impact any studies concerning the geological storage of CO_{2}. These shortcomings can be summarized as follows:
- Numerical reservoir simulation models are uncertain by nature and their effective use for the geological storage of CO_{2} require incorporation of technologies that allow such uncertainties to be quantified.
- Numerical reservoir simulation models that are constructed for non-academic, real case scenarios are very large (millions or tens of millions of cells). Therefore, these models have massive computational footprints. This fact limits their practical utilization for uncertainty analysis and quantification and for detail field studies, development planning, and optimization.
The hydrocarbon exploration and production industry have been well aware of these shortcomings and during the past several decades have spent a considerable amount of efforts to address them. The response of the hydrocarbon exploration and production industry has been to develop proxy models to address the above-mentioned issues with the numerical reservoir simulation models. The proxy models that have been developed in the E&P industry can be divided into the following major categories:
- Reduced Order Models - ROM: Numerical reservoir simulation models, especially those with formulations to model injection,
migration, and interaction of CO_{2} with the formation rock and the native fluids in the formation, include a large amount of complex
physics that have been incorporated into the simulation models. The inclusion of such complex physics sometimes requires smaller
grid blocks (cells) in order to accommodate proper convergence of the numerical techniques that are used to solve the complex,
non-linear, and higher order, partial differential equations involved in the fabric of such models.
In order to accomplish their objective of reducing the computational footprint of the numerical reservoir simulation models, the Reduced Order Models (ROM) usually incorporate one or both of the following techniques:- They tend to simplify/reduce the physics used to build the model,
- They attempt to reduce the model’s resolution in space and time. Sometimes to accomplish this, they must incorporate a reduction in the physics of the problem as well.
- Statistical Response Surfaces - SRS: Here we emphasize the word “Statistical” to draw attention to the fact that the common response surfaces development techniques in our industry use traditional statistics (not machine learning) as their base. The main differences between traditional statistics and Artificial Intelligence and Machine Learning are much deeper than some may believe and/or admit. Although these differences show themselves in the techniques used in order to accomplish their objectives, their main differences stem from their philosophical approach to problem-solving. The differences are so deep seeded that may be referenced to the dissimilarities between Aristotelian versus the Platonic view of the world.
Smart Proxy Modeling that is introduced in this book is capable of overcoming the abovementioned shortcomings of the numerical reservoir simulation models through increasing their run-time speed (shortening their computational footprints) by multiple orders of magnitude without compromising the incorporated complex physics in the numerical simulation and without decreasing its resolution in time and space.
Many reservoir engineers and reservoir modelers refer to this claim as “too good to be true”. They correctly refer to the fact that “you can’t have your cake and eat it too” and the fact that there is only so much one can do to push the envelope in creating better, more efficient algorithms or using High-Performance Computers (HPC including GPUs) to speed up the execution of traditional numerical simulation models. This is true as long as you are operating within the same paradigm that was used to develop the numerical reservoir simulation, i.e. computational science. That is the main reason behind the lack of major success in this area by those that are incorporating mathematical, statistical, and/or other approaches including multiple versions of Principle Component Analysis (PCA) and Proper Orthogonal Decomposition (POD). While these techniques have proven to enhance the speed and efficiency of the numerical simulation models, they are incapable of the type of modifications that would transform this technology to new levels of computational footprint, such as being able to perform complete, high fidelity simulation and modeling on your smartphone, tablets, or laptops in seconds.
The reason behind the success of the Smart Proxy Modeling is the fact that this technology incorporates a completely different paradigm in problem-solving. Instead of Computational Science, Smart Proxy Modeling uses Artificial Intelligence and Machine Learning to teach the detailed physics of reservoir engineering to open computer algorithms using the vast amount of data that is generated by the complex, high fidelity numerical simulation models. A good example and analogy of this technology are the driverless cars that theoretically could have been developed by numerical simulation methods but would have never been able to be practiced in real-time. It was the paradigm shift in problem-solving (a shift from computational science to data-driven analytics) that made the driverless cars a reality.