ExaSheds: Advancing Watershed System Science using Machine Learning and Data-Intensive Extreme-Scale Simulation

Abstract:

The ExaSheds project was initiated in March 2019 to explore synergies between data-driven machine learning (ML) approaches and process-based hydro-biogeochemical simulation capability with a view toward improving predictive capability for watershed function. The project is exploring the use of ML as alternatives to traditional inverse modeling and for the generation of difficult-to-observe model inputs from sparse, coarse, and indirectly related observations. Hybrid modeling approaches that use process-based simulations help guide ML to improve robustness of projections in a changing climate are also being developed. We are also adapting our integrated surface/subsurface hydrology and geochemical reactive transport codes to heterogeneous computer architectures, thus providing vastly improved simulation throughput capacity to support model-data integration and enabling large-scale simulations on leadership class computing facilities. Watershed to basin-scale projections of water availability in the Upper Colorado Water Resources Region (UCWRR) and projections of river water temperature in response to extreme weather events in the Delaware River Basin (DRB) will provide concrete use cases to drive development of an integrated thermal hydrology capability. We are exploring six essential themes: 1) ML-based preparation of model inputs, 2) ML-assisted inverse modeling, 3) hybrid ML-physics models for hydrology, 4) hybrid ML-physics models for geochemistry, 5) a simulation (hydrology and geochemistry) capability on heterogeneous computer architectures, and 6) demonstration and integration in the UCWRR and DRB.