Developing and Validating Simplified Predictive Models for CO2 Geologic Sequestration

Abstract

The objective of this study is to develop and validate a portfolio of simplified modeling approaches for CO2 sequestration in deep saline formations. It is important to ensure that the simplified models, developed using reduced physics or statistical learning-based approaches, are also capable of reproducing the full spectrum of uncertainty and sensitivity analysis results from detailed numerical simulators. A 97-run LHS design with the full-physics numerical model is used to generate the reference cumulative distribution function (CDF) of two key outcomes: plume radius and average pressure buildup in the reservoir. This is compared against CDFs from 10,000-run LHS designs with three different simplified models: (a) response surface from a Box-Behnken design and quadratic meta-model, (b) response surface from a maximin LHS design and kriging meta-model, and (c) simplified physics based methodology. The differences are quantified using two statistical measures: (1) Earth-movers distance and (2) Kolmogorov-Smirnov statistic. The statistical learning-based simplified models are found to provide a more robust representation of the reference CDF, particularly with respect to outliers.