Multi-scale hybrid modeling of terrestrial evaporation

Hybrid modeling combining physics with machine learning in recent years has pushed the frontiers of Earth science, providing an opportunity to accurately characterize traditionally elusive variables. Terrestrial evaporation (E) is one such climatic variable which couples the global water and energy cycles. Accurately estimating E is important for determining crop water requirements at the local scale, while diagnosing the vegetation state at the global scale. Despite its importance, an accurate prediction of E has proven elusive, leading to the implementation of a plethora of mechanistic and data-driven models in the last few decades. The difficulty in modeling E can be traced to the complex response of transpiration (Et; i.e., evaporation from vegetation) to various environmental stressors, which are assumed to interact linearly in global models due to our limited knowledge based on local studies. Here, we train deep learning algorithms using eddy covariance and sap flow data together with satellite observations as inputs, aiming to retrieve a universal formulation of transpiration stress (St), i.e., the reduction of Et from its theoretical maximum. Then, we embed the new St formulation within the process-based Global Land Evaporation Amsterdam Model (GLEAM). In the resulting hybrid model, the St formulation is bidirectionally coupled to the host model at the daily timescale. Comparisons against in situ data and satellite-based proxies of E demonstrate the ability of this hybrid framework to produce better estimates of St and E globally across multiple spatial scales (ranging from 1km to 0.10 degrees) (Koppa et al. 2022). The proposed framework may be extended to improve not only the modeling of E in Earth System Models but also enhance the understanding of processes which modulate this crucial climatic variable. Future work in this direction involves the development of an end-to-end hybrid model, capable of simultaneously learning and inferring St and E through differentiable programming. Our results highlight the potential of combining mechanistic modeling with machine learning, especially deep learning, for improving our understanding of complex Earth system processes which are difficult to measure directly at the scale of interest. References Koppa, A., Rains, D., Hulsman, P. et al. A deep learning-based hybrid model of global terrestrial evaporation. Nat Commun 13, 1912 (2022). https://doi.org/10.1038/s41467-022-29543-7