SCIENTIFIC MACHINE LEARNING PREDICTS NUTRIENT DYNAMICS IN EUTROPHYING LAKES

Forecasting ecosystem change is a critical task for effective management of natural resources. In lake systems, predicting nutrient concentration changes is challenging due to unobserved sources of nutrients like sewage infiltration, the complexity of interacting components, and feedback mechanisms, especially understanding the drivers of eutrophication. These factors limit the predictive power of many parametric models. Our aim is to efficiently predict unobserved nutrient inflows and provide accurate forecasts of change in algal concentrations using Scientific Machine Learning. Scientific Machine Learning combines parametric modeling (i.e., differential equations are used to describe known processes) with machine learning (i.e., data-driven predictive techniques like neural networks), to predict the quantities of interest. Specifically, we use Scientific Machine Learning to predict nutrient inputs from an unknown source, allowing us to forecast changes in the eutrophic state of a lake. This approach accurately captures the effects of environmental variability, nonlinear dynamics, system states (e.g., dissolved oxygen, nutrients, zooplankton, algal concentration), and complex interactions. Evaluating our approach with a mechanistic lake dynamics model revealed it outperforms standard parametric models in forecasting and predicting unobserved nutrient inflows. Integrating mathematical representation of known processes with machine learning enhances our ability to forecast lake ecosystem dynamics and predict Harmful Algal Blooms.

Presentation Preference: Oral

Primary Presenter: Kunal Rathore, Oregon State University (rathorek@oregonstate.edu)

Authors:

Jorge Arroyo Esquivel, California Department of Fish and Wildlife, West Sacramento, California, USA  (jorge.arroyo-esquivel@wildlife.ca.gov)

James Watson, College of Earth, Ocean and Atmospheric Sciences, Oregon State University, Corvallis, OR  (james.watson@oregonstate.edu)