A Robust Method for Adaptive Model Uncertainty Reduction, 20-R9704

Printer Friendly Version

Principal Investigators
Alexander Y. Sun
Alan P. Morris
Sitakanta Mohanty

Inclusive Dates:  04/01/07 – 03/31/09

Background - Effective management of modern oil and gas reservoirs and groundwater resources calls for adaptive decision-making frameworks facilitated by efficient data fusion methodologies and high-fidelity geologic and fluid flow models. Natural geologic systems are formed by complex geologic processes and exhibit heterogeneities at multiple scales. Consequently, model uncertainty, arising from lack of measurements and system simplification, is an inescapable aspect of geologic modeling. Recent advances in innovative in situ sensing technologies allow engineers to gather a great amount of real-time dynamic production data which, in turn, can be used to infer geologic features and reduce inaccuracies in the a priori system representation. This project was motivated by the strong need to integrate production data into reservoir and groundwater management models to support real-time or near real-time model updating and, thus, improve model reliability, enhance informed decision making, and promote the practice of adaptive management.

Approach - Sequential Bayesian updating algorithms based on Kalman filter and its variants are widely applied to update state-space models. The nonlinearity of reservoir flow models and high-dimensionality of state variables preclude the use of classic Kalman filter and its common variants. For this project, SwRI developed a grid-based local clustering ensemble Kalman filter (LCEnKF) for updating non-Gaussian parameter distributions in high-dimensional models. The algorithm combines a grid-based local ensemble analysis (localization) approach with a Gaussian Mixture Model data clustering technique to effectively reduce model dimension while mitigating filter degeneracy. Innovative algorithmic strategies were implemented to further increase computational efficiency. The LCEnKF algorithm can be parallelized easily to meet the needs of large-scale data assimilation.

Accomplishments - The effectiveness of the LCEnKF for two application areas, multi-facies fluvial aquifers and faulted reservoirs, was demonstrated. Fluvial aquifers consist of sedimentary facies with complex geospatial patterns. The lower permeability facies may trap non-aqueous phase liquids (NAPLs) and become long-term contaminant sources. This work is the first to consider updating uncertainties caused by both inner- and intra-facies heterogeneities. The distribution of a realistic geologic fault network using synthetic pumping test data also was updated. Results indicate that LCEnKF is a promising tool for assimilating high-dimensional, multimodal parameter distributions. This research resulted in two peer-reviewed journal publications, several professional conference presentations, and one grant proposal.

2009 Program Home