Physics-informed surrogate modeling for supporting climate resilience at groundwater contamination sites
Abstract
Soil and groundwater contamination is a widespread problem across the world. Contaminated sites often require decades to remediate or to monitor natural attenuation. Climate change exacerbates the problem because extreme precipitation and/or shifts in precipitation/evapotranspiration regimes can remobilize contaminants and proliferate affected groundwaters. With tools for fast and reliable contaminant plume prediction under future climate scenarios site managers and decision makers can evaluate the potential consequences and take rapid actions. Recent developments in machine learning introduce the Fourier Neural Operator (FNO) which has shown to be very effective in learning Partial Differential Equations (PDEs).
In this study, we utilize two versions of the FNO that are enhanced with U-Net architectures to model multiple resolutions: UFNO-3D and UFNO-2D. With these networks, we create surrogate flow and transport models under different CMIP5 climate scenarios. We use the Department of Energy's Savannah River Site F-Area - which has a significant groundwater contamination - as a testbed for demonstrating this capability, and evaluate the combined impact of uncertain subsurface properties and recharge rates from different climate projections. We train our UFNOs based on various loss terms that include both data-driven factors and physical boundary constraints. Results show that we can predict 1) contaminant concentration 2) hydraulic head 3) darcy's velocity from 1954 to 2100 accurately with different climate and subsurface inputs. Larger recharge rates have a complex impact on plumes with both remobilization and dilution of the contaminants. In parallel, to scale such climate resilience assessment at any site, we developed an unsupervised approach to reduce the dimensionality of the vast historical and projected climate data by identifying similar climatic regions. We leverage convolutional autoencoders combined with K-Means clustering across the United States to capture unique climate patterns from the CMIP5 model, which helps us return reliable future recharge rate projections immediately without querying large climate datasets. We hope this work can support the next level of environmental remediation modeling development under climate changes.- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFM.H45L1534W