Uncertainty Quantification of Highly-Parameterized Geothermal Reservoir Models Using Ensemble-Based Methods

Abstract

Geothermal reservoir simulations are typically time-consuming, and models made up of a few tens of thousands of model blocks can easily take hours to run. This is a considerable hindrance to model development and uncertainty quantification since current industry standard methods require running numerous simulations to invert or history-match a highly-parameterized geothermal model. Extensive uncertainty quantification using standard Markov chain Monte Carlo sampling is generally not viable in the case of a highly-parameterized geothermal model since it usually requires a significant number of simulations. Because of computational limitations in terms of time and resources, parameter uncertainty in subsurface transport models is often quantified using a limited number or a small ensemble of different numerical models. Here we explore using ensemble-based methods for approximating model parameter and predictive uncertainty in the context of models describing high-enthalpy geothermal reservoirs. First, we consider using a randomized maximum likelihood approach where we apply a recently developed randomized Levenberg-Marquardt (LM) approach to efficiently invert an ensemble of models. The randomized LM method combines adjoint code and randomized low-rank matrix factorization methods to facilitate rapid inversion of highly-parameterized models. Second, we contrast this approach with iterative ensemble smoothers, which do not require adjoint code. Both methods are well suited to high-performance parallel computing environments. Finally, we compare these two approaches with using simple local sensitivity analysis.