Search for: All records

There is an opportunity to advance both prediction accuracy and scientific discovery for phosphorus cycling in Lake Mendota (Wisconsin, USA). Twenty years of phosphorus measurements show patterns at seasonal to decadal scales, suggesting a variety of drivers control lake phosphorus dynamics. Our objectives are to produce a phosphorus budget for Lake Mendota and to accurately predict summertime epilimnetic phosphorus using a simple and adaptable modeling approach. We combined ecological knowledge with machine learning in the emerging paradigm, theory-guided data science (TGDS). A mass balance model (PROCESS) accounted for most of the observed pattern in lake phosphorus. However, inclusion of machine learning (RNN) and an ecological principle (PGRNN) to constrain its output improved summertime phosphorus predictions and accounted for long term changes missed by the mass balance model. TGDS indicated additional processes related to water temperature, thermal stratification, and long term changes in external loads are needed to improve our mass balance modeling approach.