STAR: Knowledge-Guided Machine Learning to Study Global Rivers and Nitrous Oxide

Nitrous oxide is a gas found naturally in the atmosphere that has the potential to negatively affect agricultural and ecosystem productivity. Rivers and streams are an important but poorly understood source of nitrous oxide emissions, because reactive nitrogen enters waterways and has the potential to produce nitrous oxide. Current estimates of how much nitrous oxide rivers produce vary by a factor of more than fifty, reflecting the need for an improved understanding of how nitrogen moves through freshwater systems. This project addresses this gap by producing the most accurate, data-driven estimate of riverine nitrous oxide emissions assembled. The project advances freshwater nitrogen science by developing new artificial intelligence tools, thereby contributing to a rapidly growing field with significant economic value. The project also trains American scientists in the high-demand intersection of biological science and data science, and will bring modern computational science into high school and community college classrooms across the northeastern United States. This project develops a Knowledge-Guided Machine Learning (KGML) framework to quantify riverine nitrous oxide emissions at global scale. The framework integrates a stochastic process-based model of nitrogen transformation kinetics — including nitrification, denitrification, and their dependence on oxygen, carbon, temperature, pH, and hydrology — with a gradient boosting machine learning model trained on residuals between process model predictions and observations. The project compiles the most comprehensive global database of riverine nitrous oxide concentrations (~10,000 records) and fluxes (~1,500 records) ever assembled, integrating published datasets, unpublished field measurements from six continents, and National Ecological Observation Network time series. The calibrated KGML model will be upscaled across the MERIT-Hydro global river network to generate the first spatially and temporally resolved, uncertainty-quantified global riverine nitrous oxide budget. This project leverages machine learning to resolve nonlinear biogeochemical interactions that neither empirical nor mechanistic models alone can capture. This project advances NSF’s priorities in Artificial Intelligence. 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.