Seminar:Zavud Baghirov

Studies suggest that terrestrial carbon and water cycles are strongly interconnected and should be studied as a unified system. Process-based modeling (PBM) is a key method for representing these cycles numerically, but it faces significant uncertainties due to incomplete process knowledge and limited integration of Earth observations. Machine learning (ML) offers a flexible alternative by learning from data, though it lacks physical consistency and transparency.
A promising approach, hybrid modeling, combines ML and PBM to leverage their strengths and mitigate their weaknesses. Hybrid models replace uncertain PBM parameters with ML estimations while retaining established process knowledge (e.g., mass balance), thus integrating Earth observations effectively.
This talk will first introduce the hybrid hydrological model with vegetation (H2MV), which uses dynamic neural networks to model uncertain parameters affecting water fluxes and states. The model is optimized against global observations, including terrestrial water storage anomalies (GRACE), fAPAR (MODIS), snow water equivalent (GLOBSNOW), evapotranspiration (FLUXCOM), and runoff (GRUN).
The second part will cover the hybrid hydrological carbon cycle model (H2CM), which couples a straightforward carbon cycle model to H2MV. H2CM simulates key carbon fluxes, including gross primary productivity (GPP), net primary productivity (NPP), terrestrial ecosystem respiration (TER), and net ecosystem exchange (NEE). It uses neural networks to estimate uncertain yet critical parameters such as water and carbon use efficiency. H2CM constrains GPP and NEE using observation-based data (FLUXCOM-X-BASE) and atmospheric inversion products (OCO-2 satellite inversions), indirectly constraining TER. This model outperforms state-of-the-art data-driven and process-based models in capturing NEE seasonality, especially in climatically challenging regions like the South American tropics and Southern Africa. We also show that H2CM learns interesting intermediate processes. For example, it implicitly learns the rain pulse effect—the ecosystem's short-term (daily) response to rainfall in arid regions—highlighting its capacity to generalize high-frequency processes (daily) from low-frequency constraints (monthly).