Collaborative Research: Informing River Corridor Transport Modeling by Harnessing Community Data and Physics-Aware Machine Learning

University Of New Mexico, Albuquerque NM

Investigators

Abstract

River corridors, including their adjacent and underlying sediments, are ecosystems where waters from different sources mix. This mixing controls the fate of a multitude of dissolved solutes, such as nutrients essential to the ecosystem, dissolved minerals from natural weathering, pharmaceuticals from wastewater treatment plant discharge, and contaminants from nearby sources. Practically useful computer models of how solutes are transported, including how they are exchanged back and forth between riverbed sediments and the river itself, are needed to understand water quality in rivers. Recent research suggests that these transport processes are missed by state-of-the-art computer models. This project will develop a general approach to building adaptable computer models based on recently developed tools in mathematical modeling, including artificial intelligence, to investigate how to specialize general models for particular rivers. The project will generate a large database of experimental results from river transport studies from around the globe. The database will be used to extract patterns associated with solute transport and will be disseminated broadly with the scientific community. The project team will host annual workshops to enhance database sharing, distribute educational modules on the use of artificial intelligence in hydrological sciences, and discuss approaches to standardize data collection. The goals of this project are to develop a comprehensive database of river tracer testing data for open sharing with the scientific community, and to develop and test a novel generalized model of solute transport in river corridors. The activities proposed center around the construction of a community-available, large database of tracer tests performed in streams and rivers worldwide, and its use as curricula for machine learning of model properties. Congruent data analytics will be performed to identify correlations among key variables of both river and tracer test properties, treating breakthrough curves not individually but in the tracer test sets in which they are measured. Uncertainty in experimentally measured solute concentrations will be formally addressed and used to describe model predictive power. The models selected for evaluation range from the classical transient storage model to a new model designed to address the hypothesis that residence time in the river and in the hyporheic zone both matter to exchange fluxes. Both conventional inverse modeling and machine learning tools will be applied in dual model calibration tasks, bringing uniquely powerful physics-informed neural networks to bear on this challenging problem. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.