Hybrid Modeling Improves Predictions of How Water Moves from Land to Atmosphere

Artificial intelligence guided by physics principles enables robust predictions of how ecosystems use water under extreme conditions.

Spatial distribution of daily mean evapotranspiration (ET), transpiration (T), and evaporation (E) values from the new Residual Neural Network Penman–Monteith hybrid model. — Image courtesy of Chen, H., et al., Journal of Hydrology

The Science

Water typically moves from the land to the atmosphere through soil evaporation and plant transpiration. This process is called evapotranspiration. It plays a major role in events such as drought and heat waves. However, many existing computer models struggle to distinguish between soil evaporation and plant transpiration. Researchers developed a hybrid model that provides data-informed estimates of parameters that are difficult to estimate. The model combines a physical framework with machine learning to improve predictions of evapotranspiration and its components across diverse conditions, including heat waves and drought.

The Impact

Understanding and predicting how ecosystems use water is essential to helping communities prepare for heat waves and droughts. A new hybrid model combines a form of artificial intelligence and insights from existing land surface models to create a hybrid model that improves scientists’ understanding of Earth system processes. This hybrid approach reveals how environmental factors influence evaporation and transpiration across ecosystems. These factors include radiation, soil moisture, and vegetation. These insights can help scientists improve land surface models in Earth system simulations while highlighting where traditional representations of processes break down and where methods that are physically constrained and data-informed may improve robustness.

Summary

Accurately representing evapotranspiration and its partitioning between soil evaporation and plant transpiration is a long-standing challenge in Earth system science. A multi-institutional team of researchers developed a hybrid evapotranspiration model that integrates machine learning into the Penman–Monteith modeling framework to estimate key surface conductance values while preserving physical constraints. Researchers evaluated the model using observations from 47 National Ecological Observatory Network sites spanning a wide range of environment and vegetation types. The model consistently outperformed traditional physics-based models and purely machine learning approaches, maintained accuracy at new sites, and remained stable under extreme conditions. These results demonstrate a scalable, physically grounded pathway for improving understanding of ecosystem water fluxes.

Contact

Stephen Good
Oregon State University
stephen.good@oregonstate.edu

Rich Fiorella
Los Alamos National Laboratory
rfiorella@lanl.gov

Funding

This work was primarily supported by the Environmental System Science program within the U.S. Department of Energy Office of Science’s Biological and Environmental Research program. Support was also received from the Zegar Family Foundation and the National Science Foundation Division of Environmental Biology.