Journal Article

Hybrid‐Modeling of Land‐Atmosphere Fluxes Using Integrated Machine Learning in the ICON‐ESM Modeling Framework

Carvalhais, Nuno, ElGhawi, Reda, Reimers, Christian, Körner, Marco, Schnur, Reiner, Reichstein, Markus, Winkler, Alexander J.

2025

Overview

The water and carbon exchange between the land surface and the atmosphere is regulated by meteorological conditions and plant physiological processes. Traditional mechanistic modeling approaches, for example, the Earth system model ICON‐ESM with the land component JSBACH4, are hampered by relatively rigid parameterizations for stomatal conductance to model land‐atmosphere coupling. We develop a hybrid modeling approach integrating data‐driven flexible parameterizations based on eddy‐covariance flux measurements (FLUXNET) with mechanistic modeling. We replace specific empirical parametrizations of the coupled photosynthesis (gross primary production [GPP]) and transpiration (Etr${E}_{\\text{tr}}$ ) modules with feed‐forward neural network models pre‐trained on observations. In a proof‐of‐concept, we demonstrate that our approach reconstructs original JSBACH4 parameterizations for stomatal conductance (gs${g}_{s}$ ), maximum carboxylation rates (Vcmax${V}_{\\text{cmax}}$ ) and the maximum electron transport rates (Jmax${J}_{\\text{max}}$ ), that decisively control GPP and Etr${E}_{\\text{tr}}$ . We then replace JSBACH4's original parametrizations by calling the emulator parameterizations trained on original JSBACH4 output using a Python‐Fortran bridge. Adapting the approach to observational data, Hybrid‐JSBACH4 infers these parametrizations from eddy‐covariance measurements to construct observation‐informed modeling of water and carbon fluxes in JSBACH4. The mean hourly residuals of Etr${E}_{\\text{tr}}$in Hybrid‐JSBACH4 with respect to FLUXNET observations vary between −0.1 and 0.15 kg m−2 hr−1 while the JSBACH4 Etr${E}_{\\text{tr}}$residuals vary between −0.3 and 0.2 kg m−2 hr−1 for forest and grassland sites. The mean hourly residuals for GPP of Hybrid‐JSBACH4 with respect to observations vary between −0.5 and 0.5 gC m−2 hr−1, compared to the original JSBACH4 with residuals ranging between −1.0 and 0.5 gC m−2 hr−1, for forest and grassland sites. Our Hybrid‐JSBACH4 model improves the representation of plant physiological responses, and reduces biases in transpiration and GPP simulations under varying atmospheric dryness and water availability conditions. Plain Language Summary This study presents a novel hybrid modeling approach, Hybrid‐JSBACH4, that combines machine learning with traditional process‐based modeling to enhance the representation of carbon and water exchanges in terrestrial ecosystems. Specifically, we integrate a feed‐forward neural network (FNN) trained on eddy‐covariance flux data from FLUXNET to replace key parameterizations in the JSBACH4 model, which is part of the ICON‐ESM framework. We demonstrate that our hybrid model captures latent variables, stomatal conductance (gs${g}_{s}$ ), maximum carboxylation rates (Vcmax${V}_{\\text{cmax}}$ ), and maximum electron transport rates (Jmax${J}_{\\text{max}}$ ) and emulates the gross primary productivity gross primary production (GPP) and transpiration (Etr${E}_{\\text{tr}}$ ) outputs of the original JSBACH4 model. By utilizing FLUXNET observations, the Hybrid‐JSBACH4 offers a more adaptable and data‐driven representation of plant physiological responses for GPP and Etr${E}_{\\text{tr}}$ . The results indicate that Hybrid‐JSBACH4 significantly improve simulations of GPP and Etr${E}_{\\text{tr}}$under varying environmental conditions, effectively capturing the influences of vapor pressure deficit and soil water content. Compared to JSBACH4, Hybrid‐JSBACH4 reduces biases in transpiration and GPP, better‐capturing ecosystem dynamics. These refinements enhance our ability to predict ecosystem responses, particularly in balancing supply‐driven and demand‐driven water stress, leading to a more physically consistent and unbiased representation of land‐atmosphere interactions. Key Points The Hybrid‐JSBACH4 model integrates pre‐trained neural networks with land surface modeling to improve photosynthesis and transpiration representation The Hybrid‐JSBACH4 reduces transpiration and water‐use efficiency biases under varying soil moisture and atmospheric demand conditions Structural rigidity in carbon cycle formulation limits photosynthesis improvement in trade‐off for improved transpiration Hybrid‐JSBACH4

• Publisher: John Wiley & Sons, Inc,American Geophysical Union (AGU)
• Subject: Atmosphere / Atmospheric boundary layer / Carbon / Carbon exchange / carbon‐water fluxes / Carboxylation / Climate change / Data assimilation / Earth system models / Ecosystems / Environmental Chemistry / General Earth and Planetary Sciences / Global and Planetary Change / Grasslands / Hybrid modeling / Hydrologic cycle / Hydrology / ICON‐ESM / JSBACH4 / Machine learning / Meteorological conditions / Modelling / Neural networks / Photosynthesis / Physics / Physiological responses / Physiology / Primary production / Stomatal conductance / Transpiration / Water / Water availability / Water shortages