Keywords
Carbon Sequestration, Metamodel, Proxy Model, Response Surface Model
Location
Session A2: Sharing Scientific Environmental Data and Models
Start Date
17-6-2014 2:00 PM
End Date
17-6-2014 3:20 PM
Abstract
Full-physics models for geologic carbon sequestration applications can be time-consuming to run. As a result, optimization of a particular response over a range of predictor values can be difficult. A common approach is to use a small number of simulation runs to develop a metamodel that can approximate the system response with much less computing time; this metamodel can then be used for optimization. Choosing a metamodeling approach is not always straightforward, and it can also be difficult to determine how well a model is fitting after it is trained. Here, we present a case study for a CO2 injection problem in the ARCHES province in the American Midwest, wherein a cross-validation approach was used to evaluate the quality of several metamodel fits.
To begin, the STOMP-CO2 simulator was used to generate 36 simulation runs by varying the values of three predictors. The modeling of CO2 injection into a closed volume results in three responses describing the amount and extent of the CO2 stored in the model domain, as well as the pressure at the injection site. These 36 runs were used to train several metamodels, which included quadratic polynomial regression, kriging, MARS, AVAS, and TPS. Cross-validation was then used to evaluate the root mean squared error (RMSE) of these metamodels over the design space. In all cases, the kriging model had the most favorable RMSE.
Included in
Civil Engineering Commons, Data Storage Systems Commons, Environmental Engineering Commons, Hydraulic Engineering Commons, Other Civil and Environmental Engineering Commons
Evaluation of Metamodeling Techniques on a CO2 Injection Simulation Study
Session A2: Sharing Scientific Environmental Data and Models
Full-physics models for geologic carbon sequestration applications can be time-consuming to run. As a result, optimization of a particular response over a range of predictor values can be difficult. A common approach is to use a small number of simulation runs to develop a metamodel that can approximate the system response with much less computing time; this metamodel can then be used for optimization. Choosing a metamodeling approach is not always straightforward, and it can also be difficult to determine how well a model is fitting after it is trained. Here, we present a case study for a CO2 injection problem in the ARCHES province in the American Midwest, wherein a cross-validation approach was used to evaluate the quality of several metamodel fits.
To begin, the STOMP-CO2 simulator was used to generate 36 simulation runs by varying the values of three predictors. The modeling of CO2 injection into a closed volume results in three responses describing the amount and extent of the CO2 stored in the model domain, as well as the pressure at the injection site. These 36 runs were used to train several metamodels, which included quadratic polynomial regression, kriging, MARS, AVAS, and TPS. Cross-validation was then used to evaluate the root mean squared error (RMSE) of these metamodels over the design space. In all cases, the kriging model had the most favorable RMSE.
