Speaker
Description
Dissolution of solid mineral in porous media due to the introduction of reactive fluids is of utmost importance for a wide range of subsurface applications, including CO2 storage, geothermal systems, hydrogen technology, and enhanced oil recovery. The conditions of the injection process as well as the mineral properties strongly influence the resulting dissolution pattern, leading to compact, uniform, wormholing, or channelling dissolution that change the permeability and flow properties of the reservoir. Direct numerical simulation of the pore-scale dissolution process is difficult, with many thousands of CPU hours required for even relatively small regions of pore-space, making routine prediction of realistic volumes relevant to subsurface applications impractical. Deep learning has the potential to revolutionise this approach, both by increasing the speed of the solver and providing upscaled models for accurate modelling of dissolution in large domains.
In this work we leverage our fast, efficient dissolution numerical model in our open-source toolbox GeoChemFoam to run 2D simulations of dissolution on ultra-large synthetic, stochastically created geometries with varying levels of pore-space heterogeneity, flow, and reaction rates. We then use these numerical results as a training dataset for two deep learning models. (1) Using image analysis on subsections of the model results we extract flow and reactive parameters and train a deep neural network to predict the porosity and permeability changes on Darcy-scale grids. (2) We develop efficient deep learning emulators for geochemical reactions using deep residual recurrent neural network to develop highly predictive reduced order models using limited training data and utilizing U-net architectures to perform approximate explicit time stepping for the dynamical system. Both trained deep learning models are then integrated with GeoChemFoam’s solvers for increased speed and upscaling capability.
| Country | UK |
|---|---|
| Acceptance of the Terms & Conditions | Click here to agree |
