AGU Fall Meeting 2023

B34B-02 Hybrid-Modeling of Land-Atmosphere Fluxes Using Integrated Machine Learning in the ICON-ESM Modeling Framework

Thursday, 14 December 2023, 01:10-01:20

Reda El Ghawi, Christian Reimers, Markus Reichstein, Marco Körner, Nuno Carvalhais and Alexander J. Winkler

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. Traditional mechanistic modeling approaches, e.g., the ICON-ESM with the JSBACH4 model, have long been used to study the land-atmosphere coupling. However, their ability to generalize is hampered by relatively rigid functional representations of terrestrial biosphere processes, e.g. semi-empirical parametrizations of stomatal conductance. Here, we develop data-driven, flexible parameterizations controlling the terrestrial processes based on eddy-covariance flux measurements using machine learning (ML). Specifically, we introduce a hybrid modeling approach (integration of data-driven and mechanistic modeling), Hybrid-JSBACH, that aims to replace the empirical parametrizations of the coupled photosynthesis (GPP) and transpiration (E_tr) modules with pre-trained ML models. We do so by constructing a Python-FORTRAN bridge, allowing us to dynamically call the pre-trained data-driven parameterizations based on JSBACH4 output to replace JSBACH4’s original parameterizations for stomatal (g_s) and aerodynamic (g_a) conductance, the carboxylation rate of RuBisCO (V_cmax), the photosynthetic electron transport rate for RuBisCO regeneration (J_max), and internal leaf CO₂ partial pressure (c_i) with the Neural Network model.

Based on observational records at six eddy-covariance sites, three forest and three grassland sites, we not only are able to predict water and carbon fluxes but also infer latent dynamic features, i.e., g_s, g_a,V_cmax, J_max, and c_i.

This modeling framework will then serve as the foundation for the coupled simulation of the land-atmosphere interface in ICON-ESM, where the biospheric processes are represented by our hybrid land-surface model.

B34B-08 Improving the Representation of Terrestrial Ecosystem Processes via Machine Learning Approaches (Invited)

Thursday, 14 December 2023, 02:10-02:20

Nuno Carvalhais

Quantifying the responses of terrestrial carbon and water cycles to changes in climate is central to address the challenges of global warming and to increasing atmospheric CO2. Much of the current understanding and quantification of short-to-long term ecosystem responses emerges from the model integration of ecophysiological understanding and ecological knowledge. When models and observations are brought together, the limitations in process representation and in model generalization become apparent, as also evident in several generations of coupled model experiments. Here, we bring forward machine learning approaches to improve the parameterization, generalization and representation of terrestrial carbon and water cycles in process-based models. Leveraging in situ and Earth observation data, we develop hybrid modeling approaches to improve the representation of ecosystem processes from local to regional and scales. Namely, we showcase how coupling process-based with deep learning models overcomes equifinality and improves generalization in representing photosynthesis, as well as the representation of temporal dynamics and spatial patterns in land atmosphere fluxes from fire emissions. We further leverage on explainable AI for interpreting the structure of the emerging deep learning models. Ultimately, we argue for methodological diversity and observational heterogeneity to further improve the modeling and understanding of the climate responses of terrestrial ecosystems, and reduce their uncertainties.