Physics‐Constrained Machine Learning of Evapotranspiration

Estimating ecosystem evapotranspiration (ET) is important to understanding the global water cycle and to study land‐atmosphere interactions. We developed a physics constrained machine learning (ML) model (hybrid model) to estimate latent heat flux (LE), which conserves the surface energy budget. By comparing model predictions with observations at 82 eddy covariance tower sites, our hybrid model shows similar performance to the pure ML model in terms of mean metrics (e.g., mean absolute percent errors) but, importantly, the hybrid model conserves the surface energy balance, while the pure ML model does not. A second key result is that the hybrid model extrapolates much better than the pure ML model, emphasizing the benefits of combining physics with ML for increased generalizations. The hybrid model allows inferring the structural dependence of ET and surface resistance (rs), and we find that vegetation height and soil moisture are the main regulators of ET and rs. A physics constrained machine learning model is developed using the FLUXNET2015 Tier 1 data set. This new approach is able to effectively retrieve latent heat flux while constraining energy conservation in the surface energy budget. This hybrid model has better performance in extrapolation than a pure machine learning model. A physics‐constrained machine learning model of evapotranspiration (hybrid model) is developed and trained using the FLUXNET 2015 data setThe evapotranspiration retrieved by the hybrid model is as accurate as pure machine learning model and also conserves surface energy balanceThe hybrid model better reproduces extremes and thus better extrapolates compared to the pure machine learning approach