Free Event


CO2 flooding is widely used as Enhanced Oil Recovery (CO2-EOR) in conventional reservoirs, garnering special interest as a method of Carbon Capture Utilization and Storage (CCUS) to reduce carbon footprint. Despite accumulated experience and knowledge regarding the CO2 process from various projects, each new project necessitates tedious simulation studies to estimate recoveries and profitability. Reservoir simulation studies offer detailed technical insights, with validity tied to input data accuracy. The time-intensive nature of reservoir simulation renders it ineffective for initial screening among multiple candidates.

In this study, we introduce an integrated approach to estimate oil recovery and resulting CO2 storage potential promptly and effectively using AI-generated Proxy Model. This model encompasses reservoir characteristics, fluid properties, rock and fluid interactions, and operational constraints, predicting the performance profile of new projects, as long as their values fall within the ranges of the established proxy model. The Proxy Model was constructed based on data from over 2000 numerical simulation cases, covering a wide range of CO2-EOR pilot and mature projects operated worldwide. Specific templates for Relative Permeability Curves and Reservoir Fluids Compositions were introduced, while Reservoir Porosity, Permeability, and Initial Reservoir Pressure were generated as samples using Latin Hyper Cube experiments within ranges inspired by collected real field data.

The AI Proxy model was built using an Artificial Neural Network (ANN). Despite the time advantage offered by AI models, concerns exist regarding their "black box" nature, which generates output data from input data without explicit governing equations reflecting physical processes. Physical compliance quality control measures were thus implemented to ensure the ANN model accurately captured the physics of the system and predicted outputs with acceptable accuracy. Several checkpoints were completed to ensure result reliability, including model validation using a separate dataset from training data, resulting in nearly perfect matches between ANN model predictions and actual results. Performance evaluation metrics demonstrated good overall performance, and sensitivity checks were conducted to verify expected behavior.

Reservoir response and behavior plotted as dimensionless rate and derivative profiles versus dimensionless time, generating type curves for various reservoir parameters. This approach serves as a valuable first screening tool for potential CO2 EOR projects.


0 people are interested in this event

The University of Kansas Powered by the Localist Community Event Platform © All rights reserved