Mixed-dimensional model assimilation for aquifer characterization and CO2 plume monitoring

Abstract:

CO₂ storage monitoring and verification (M&V) requires robust models which can demonstrate compliance with field-observations. Examples include 3D static geological models, 3D dynamic models and 2D vertically-integrated models, which exist at differing spatial scales and dimensionality. Ensuring that these models are consistent is essential in robust data integration, and in the prediction of compliance metrics such as CO₂ plume migration and regional pressurisation.

Here we present a generalised upscaling-downscaling computational approach to assimilate 3D static/dynamic models with 2D depth averaged flow models, which are constrained with complementary field-information. We demonstrate this workflow in the context of the Paaratte aquifer in the Otway Basin, Australia, which forms the storage unit for the CO2CRC Otway Stage 3 project. Five new wells have been completed at ~1.5km depth in the formation, and form the basis of the M&V effort.

The 3D geological model is built on well-log, core-data and seismic interpretation; it is constrained by multi-point statistics, depositional facies variograms, and acoustic impedance and permeability/porosity distributions at sub-metre scale. The 2D depth averaged flow model is generated from inversion of cross-well pressure tests. Utilising a Bayesian adjoint inversion method, it generates large-scale (>50m) spatially varying diffusivity and porosity-thickness maps of the aquifer, matching the pressure data. Upon repeat pressure tests after CO₂ has been injected, the method also generates a probability CO₂ plume map, which can be compared to 3D dynamic models and time-lapse seismic.

We use well constrained formation height and upscaled well-log data to inform the Bayesian prior in the flow model inversion. The inversion then generates updated depth-averaged diffusivity and porosity-thickness fields based on new pressure information. The depth-averaged model can then be downscaled back to the 3D model through solution of an inverse-upscaling optimisation problem stabilised by a kriged spatial bias. Iteration of this process results in a robust set of aquifer models; the large-scale features obtained from the depth-averaged model are retained alongside the well-known features of the 3D model, leading to more accurate compliance predictions.