|
|
| |
Application of Neural Operators for Reactive Transport Modeling of CO2 Reinjection Into Geothermal Reservoirs
Tao WANG, James W. PATTERSON, David J. BYRNE, Hewei TANG
[The University of Texas at Austin, USA]
Emission of dissolved non-condensable CO2 is a major contributor to the geothermal carbon footprint of conventional geothermal energy development, and reinjection of CO2 with produced brine is a technique to reduce carbon emission intensities. A previous study has found that the addition of CO2 can significantly reduce the rate of silica precipitation in greywacke (a typical Taupo Volcanic Zone reservoir rock) sand packs. Silica precipitation is responsible for the injectivity decline of conventional geothermal wells. In this study, we propose to leverage neural operators to model geothermal reactive transport. Neural operators are designed to learn the solution operators of partial differential equations (PDEs) by mapping between infinite-dimensional function spaces. Fourier neural operators (FNOs) have good potential to approximate complex PDE systems using the Fourier transform. Meanwhile, the integration of trained neural operator surrogate models into an inversion workflow will significantly speed up the process. In this study, FNO surrogate models are trained to model reactive transport of CO2 reinjection column experiments and to replace a conventional reactive transport numerical solver (PFLOTRAN) in a PESTPP-IES inversion workflow for estimation of geochemical parameters. The trained FNO surrogates can predict the temporal outlet silica concentrations, calcium concentrations, and pH values of the complex reactive transport system with high accuracy (R2 scores up to 0.9998) and superior time efficiency (95.6% faster than PFLOTRAN). With the implementation of FNO surrogates, the time cost of the inversion workflow has reduced by 34.5%, while achieving almost identical results as the workflow with PFLOTRAN. The investigation demonstrated the speedup advantages of implementing FNO surrogates against conventional reactive transport solvers in inverse modeling. The study will be instructive to employ neural operators to model reactive transport and estimate reactive transport parameters in geothermal applications.
Topic: Modeling