Using Physics-Informed Machine Learning to Partition Eddy Covariance based Evapotranspiration

Due to the coupling of gas exchange in terrestrial ecosystems, transpiration (T) estimates are a key insight into global climate, water, and carbon patterns. While extensive datasets of evapotranspiration (ET) are abundant, largely due to global eddy covariance networks, independent measurements of T are relatively sparse. As such, there is a need to partition T out of eddy covariance-measured ET so that we can better understand how the individual components of the terrestrial water vapor flux are contributing to the global water cycle and changing under a warming climate. Physics-informed machine learning (PI-ML) presents a novel way to partition eddy covariance-measured ET even without extensive T datasets for validation. PI-ML works to constrain the model to obey underlying governing equations driving the system dynamics, giving unique insights into model estimates. Here, PI-ML is introduced to estimate the ecosystem transpiration ratio (T/ET) using eddy covariance data collected from a soybean field in Ontario, Canada. The model, founded on the principle that transpiration follows a sine curve over a 24hr period, constrains both the upper and lower bounds of T/ET by assuming transpiration is negligible overnight and that the agricultural site has periods with negligible soil evaporation. The PI-ML model was validated against leaf-level transpiration measurements collected over the 2025 growing season as well as compared to results from other eddy covariance-based ET partitioning methods.  Preliminary results for the 2019 growing season showed the PI-ML estimated a daytime average T/ET of 0.584 in the soybean field, compared to 0.468 estimated from an underlying water use efficiency-based method and 0.653 estimated from a method using data-driven machine learning. The physically realistic T estimates produced by the PI-ML show the model’s ability to accurately represent the ecosystem dynamics of the soybean site where it was applied. Accurate T estimates give way for better management of our limited water resources, leading to increased water quality and food security when used in agricultural settings.