Advancing Carbon Sequestration through Smart Proxy Modeling: Leveraging Domain Expertise and Machine Learning for Efficient Reservoir Simulation
Abstract
ABSTRACT
Advancing Carbon Sequestration through Smart Proxy Modeling: Leveraging Domain Expertise and Machine Learning for Efficient Reservoir Simulation
Iman Oraki Kohshour
Geological carbon sequestration (GCS) offers a promising solution to effectively manage extra carbon, mitigating the impact of climate change. This doctoral research introduces a cutting-edge Smart Proxy Modeling-based framework, integrating artificial neural networks (ANNs) and domain expertise, to re-engineer and empower numerical reservoir simulation for efficient modeling of CO2 sequestration.
In the absence of sufficient subsurface CO2 sequestration dara, by utilizing data from numerical reservoir simulation, this research creates realistic and yet complex data sets models at a resolution of each grid-cell, specifically tailored to research and applied objectives, facilitating the exploration of various operational scenarios and dynamic conditions.
The ANNs, trained on extensive reservoir simulation data, accurately replicating fluid flow behavior, resulting in significant computational savings compared to traditional numerical simulation models. The approach addresses several types of static and dynamic data augmentation ranging from spatial complexity, spatio-temporal complexity, and heterogeneity of reservoir characteristics. By incorporating extensive domain-expertise-based feature generation, this framework honors precise representation of reservoir overcoming computational challenges associated with numerical reservoir tools.
The results showed that all the ML models achieved very good accuracies and high efficiency. In addition, we also ranked the feature importance of the data in the CO2 saturation estimation models using different algorithm, and based on the results, quality of the path between the focal cell and injection wells was found to be the most significant factor in CO2 saturation and pressure estimation models.
These insights significantly contribute to our understanding of CO2 plume monitoring, paving the way for breakthroughs in investigating reservoir behavior at a minimal computational cost. The transformative nature of this research has vast implications for advancing carbon storage modeling technologies. By addressing the limitations of traditional numerical reservoir models and harnessing the synergy between machine learning and domain expertise, this work provides a practical workflow for efficient decision-making in sequestration projects. The ultimate objective is to establish that conformance can be achieved by demonstrating predictive and replicative capabilities of smart proxy modeling.