Parameter learning and scaling in hybrid ecosystem models for improved understanding of carbon and water dynamics

Accurate modeling of terrestrial carbon, energy, and water cycles is critical for understanding ecosystem processes and their responses to environmental change. However, a key challenge lies in the parameterization of these complex processes, which vary across scales and ecosystems. This study investigates how hybrid modeling approaches can enhance ecosystem parameter learning and provide deeper insights into terrestrial carbon and water dynamics across Europe. Specifically, we used a hybrid modeling framework that integrates the coupled photosynthesis-evapotranspiration model as a differentiable ecosystem model with a deep neural network to optimize parameter learning. Long-term observations from multiple FLUXNET sites across Europe, including daily evapotranspiration (ET) and gross primary productivity (GPP) data, were used to constrain model parameters in an end-to-end mode. The calibrated model was then used to generate spatial distribution maps of key ecosystem parameters, revealing how they vary under different climatic and ecological conditions. 

Results demonstrate that the hybrid model significantly improves simulation accuracy for ET and GPP while capturing parameter variability across European ecosystems. Post-hoc analyses of the embedded neural network quantified the influence of key environmental drivers, such as climate, soil properties, and vegetation, on the learned parameters. These results highlight the value of hybrid modeling for improving understanding of ecosystem processes, providing actionable insights for climate adaptation and ecosystem management in Europe and for improving terrestrial biosphere models.