Hybrid-Modeling of Land-Atmosphere Fluxes Using Machine Learning integrated in the ICON-ESM Modeling Framework

The exchange of water and carbon between the land-surface and the atmosphere is regulated by meteorological conditions as well as plant physiological processes. Accurate modeling of the coupled system is not only crucial for understanding local feedback loops, but also for global scale carbon and water cycle interactions. Mechanistic modeling approaches, e.g., the Earth system model ICON-ESM with the land component JSBACH4, are mainly applied to study land-atmosphere coupling. However, these models are hampered by relatively rigid and ad-hoc formulations of terrestrial biospheric processes, e.g., semi-empirical parametrizations for stomatal conductance, which often result in non-plausible and biased dynamics.

Here, we develop data-driven, flexible parametrizations controlling terrestrial carbon-water coupling based on eddy-covariance flux measurements (FLUXNET) to be implemented in the JSBACH4 model. Specifically, we introduce a hybrid modeling approach (integration of data-driven and mechanistic modeling), that aims to replace specific empirical parametrizations in JSBACH4’s modules computing coupled photosynthesis (gross primary production, GPP ) and transpiration (E_tr) fluxes based on a multi-task feed-forward neural network (FNN) modelling approach pre-trained on observations. First, as a proof-of-concept, we train parametrizations based on original JSBACH4 output to showcase that our approach succeeds in reconstructing the original parametrizations, namely latent dynamic features for stomatal conductance (g_s), the maximum carboxylation rate (V_cmax) and maximum electron transport rate (J_max) that decisively control GPP and E_tr . Second, we replace JSBACH4’s original parametrizations by dynamically calling the emulator parameterizations trained on original JSBACH4 output using a Python-FORTRAN bridge. This allows us to understand how small changes can propagate over time and enables us to evaluate the effects of data-driven parameterizations on the results produced by the coupled land-surface model. In the last step, we adopt the approach to infer these parametrizations from FLUXNET observations to construct an observation-informed modelling of water and carbon fluxes within the land model JSBACH4.

Our hybrid approach almost perfectly reproduces the original JSBACH4 parametrizations by emulating the latent variables yielding R ² values ranging between 0.99-1.0 for GPP and E_tr  at hourly scale for forest and grassland sites. JSBACH4 equipped with these plugged-in emulations of the parametrizations reveal that the NN parametrizations are capable of reproducing the targets with relatively high accuracy while learning g_s , V_cmax and J_maxwithout prior information. By training the hybrid model on FLUXNET observations and we obtain observations-informed parametrizations to be plugged-in JSBACH4. We find that Hybrid-JSBACH can better capture the variability of GPP and E_tr  across different ranges of atmospheric and soil dryness in comparison to JSBACH by analyzing the mean hourly residuals for the target variables. While challenges persist in fully integrating carbon and water cycles due to physical constraints in carbon cycle modeling, the Hybrid-JSBACH modeling framework already enables observation-guided coupling of land-atmosphere interactions for the water cycle with key biospheric processes represented by our hybrid observation-informed land-surface model. These developments are key to critically advance our understanding of hydrological processes and linked feedbacks in the climate system, especially in the context of changing climatic conditions.