Seminar:Reda El Ghawi

Accurately capturing interactions between the terrestrial biosphere and the atmosphere is critical for understanding the feedbacks that regulate the Earth system. While traditional Earth System Models (ESMs) offer mechanistic rigor, they are often hampered by rigid, semi-empirical parameterizations. Conversely, purely data-driven Deep Learning (DL) models excel at capturing complex patterns but frequently lack physical consistency and fail to extrapolate to novel climate states. This research demonstrates the potential of hybrid modeling—integrating mechanistic knowledge directly into DL frameworks—to overcome these limitations. The Hybrid-JSBACH4 framework represents a step towards integrating ML-learnt parametrizations into the JSBACH4 land surface model within the ICON-ESM framework using a Python-Fortran bridge. This integration significantly reduced biases in transpiration and water-use efficiency under high-stress conditions like drought and high atmospheric dryness at localized FLUXNET sites. Moving from uncoupled site-specific models towards global coupled simulations allows us to better understand and capture emergent feedbacks, such as soil moisture-precipitation coupling and vegetation’s role in modulating boundary-layer development. Preliminary results show that the hybrid model predicts transpiration with mean global R² = 0.54 while JSBACH has a mean global R² =0.42 with respect to transpiration from FLUXCOM. This framework will be transferred into ICON-ESM, for fully coupled ML- and Hybrid-learnt parametrizations for atmosphere and land to be jointly tuned.