Machine Learning-Assisted Multifidelity Biogeochemical Modeling for Watersheds and River Basins
Abstract
A full treatment of water flow and biogeochemical reactive transport at watershed to river basin scales is not currently possible, and yet there is a critical need to provide improved estimates of fluxes of nutrients and contaminants for both scientific and water management purposes at these scales. The use of hybrid multi-fidelity predictive models that take advantage of machine learning (ML) techniques offers an attractive option to overcome the obstacles associated with computational expense, especially insofar as it is possible to maintain process fidelity for heterogeneously distributed biogeochemical processes interacting with the hydrological cycle. ML can play an important role in at least four ways: 1) ML can facilitate the inclusion of diverse big data in physics-based models for water and biogeochemistry through downscaling and upscaling approaches, 2) ML can enable the development of reduced order/dimension and surrogate models that capture watershed and river basin function with reduced computational expense, 3) On The Fly ML can be used for automation of uncertainty quantification (UQ) to choose dynamically the level of fidelity and computational expense that is adequate for a given river basin-scale simulation, and 4) On Demand ML (ODML) can be used to gradually reduce the number of full predictive calculations that are needed to describe the watershed to river basin-scale biogeochemical function, essentially replacing full physics-based simulation with continuously improving surrogate models (Leal et al 2020). The ODML model begins with zero knowledge at the beginning of the simulation and then gradually learns key biogeochemical transport calculations using full (3D) or reduced dimension (1D flowtubes or 2D cross sections) simulations. These key calculations are then used as often as possible to predict similar calculations. ML can provide a uncertainty quantification (UQ) based approach relying on selective comparison with observational data and high resolution physics-based simulations to automatically choose the fidelity of a biogeochemical approach (e.g., high resolution RTM versus 1D flowtube versus surrogate) to balance the demands of computational efficiency and process fidelity.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFM.H31B..05S