H071-04
A Hybrid Modeling Framework Combining Machine Learning and Third-wave AI with Process-resolving Simulations to Predict River Water Quality
A Hybrid Modeling Framework Combining Machine Learning and Third-wave AI with Process-resolving Simulations to Predict River Water Quality
Wednesday, 9 December 2020: 16:12
Virtual
Abstract:
Land water interfaces (LWI) are active zones for biogeochemical transformation and interfacial fluxes. LWI dynamics are therefore known to impact subsurface geochemical exports and river water quality. To accurately capture LWI dynamics, watershed scale-reactive transport models are needed. However, such models are computationally demanding and not easy to parametrize. In this study, we present a hybrid modeling strategy combining machine learning (ML) and third-wave AI with process-resolving simulations that will capture LWI dynamics and predict river water quality with both process fidelity and computational tractability. Specific objectives of this study include (a) quantifying hydrologic exchange and biogeochemical transformations at LWIs and (b) computing subsurface geochemical exports and river water quality as a function of the characteristic watershed features (e.g., topography, wetness index, meanders, sinuosity, and amplitudes). We conducted this study at the East River Watershed, located in Colorado, which is a study site of Berkeley Lab’s Watershed Function Scientific Focus Area. To quantify LWI dynamics, we carried out three-dimensional reactive flow and transport simulations for a 10-meander system in the upper East River Watershed using PFLOTRAN, an open-source, high-resolution, three-dimensional, reactive flow and transport code. To account for interfacial fluxes and concurrent biogeochemical transformation, we integrated RiverFlotran, a river chemistry module, with PFLOTRAN. The simulation results indicated that river stage, bathymetry, and meander geometry influence hydrologic exchanges at the LWI. The results further demonstrated that higher hydrologic exchange produces hot spots of nitrogen and other redox species at the LWIs. Finally, we develop a scaling relationship between hydrologic exchange and biogeochemical transformations with river landform features characterized by various sinuosity and amplitude of meanders, topography, and residence times. This scaling relationship is used to predict subsurface geochemical exports and downstream river water quality of the more extensive East River system.