2026-08-12 –, Room 3
Scientific modeling often forces a choice between flexible but opaque neural networks and interpretable process-based models that can be too rigid for real-world data. Hybrid modeling bridges this gap by combining mechanistic structure with machine-learning flexibility. In this talk we introduce EasyHybrid.jl, a user-friendly Julia package that makes hybrid modeling accessible across disciplines. We demonstrate the approach on a concrete problem: partitioning eddy-covariance net carbon fluxes into photosynthesis (a CO₂ sink) and ecosystem respiration (a CO₂ source), while estimating how strongly respiration responds to temperature. Temperature sensitivity is summarized by Q10, the factor by which respiration changes for a 10 K warming (e.g., Q10 = 2 means doubles per 10 K). We present cross-site results across hundreds of FLUXNET eddy covariance towers and show that, even when inferred jointly with hybrid flux partitioning, the learned temperature sensitivity exhibits a relatively narrow convergence across ecosystems.
The talk shows how to use EasyHybrid.jl to define a process-based model, turn it into a hybrid model, and train it end-to-end on eddy-covariance data. We start from typical tabular data (meteorology and fluxes) and show how EasyHybrid.jl connects predictors, mechanistic forcings, and targets using named variables, while relying on Lux.jl for the neural components.
We then build a hybrid flux-partitioning model in which ecosystem respiration follows the mechanistic Q10 temperature response, but the base respiration term and radiation-use efficiency for photosynthesis are learned as functions of environmental drivers. The key design choice is explicit: which parameters are learned globally (one constant temperature-sensitivity parameter Q10), which are fixed, and which vary in time and/or space through neural networks. Training uses standard optimizers such as Adam or RMSProp.
In an outlook we explore an LSTM-based variant of the hybrid model. Instead of predicting base respiration from instantaneous drivers only, the LSTM takes sequences of past environmental drivers to test whether ecosystem memory changes flux partitioning and the inferred temperature sensitivity.
The research presented in this talk is the result of joint efforts by Lázaro Alonso, Kilian Hochholzer, Laura van der Poel, Lukas Schirren, Andrés Tangarife-Escobar, and Ritesh Moon. EasyHybrid.jl was started by Markus Reichstein. We thank Jake Nelson and the FLUXCOM team for providing access to curated FLUXNET data.
Bernhard Ahrens leads the “Modeling Interactions in Soil Systems” group at the Max Planck Institute for Biogeochemistry in Jena and works in the Biogeochemical Integration Department under Prof. Dr. Markus Reichstein. He is a geoecologist and global change ecologist working at the interface of process-based soil organic matter modeling, data integration, and machine learning. Methodologically, he codevelops the software package EasyHybrid.jl to embed neural networks into process-based models. He (co-)developed the soil organic matter turnover models COMISSION v1.0 and v2.0 as well as the Jena Soil Model. He supervises several PhD and postdoctoral projects, including within the AI4SoilHealth and WETSCAPES2.0 consortia.
Ritesh is a PhD researcher working at the intersection of hydrology and machine learning, developing physically informed machine learning (PIML) frameworks that integrate process-based hydrological models with deep learning to improve streamflow prediction in complex and regulated catchments. Using large-sample datasets such as CAMELS-GB, he incorporates hydrological signatures and process-based insights to enhance both predictive accuracy and model interpretability. His work focuses on bridging physics-based understanding with modern AI to build robust, scalable, and transparent tools for water resource management.
A scientist at the Max Planck Institute for Biogeochemistry, advancing earth system models through hybrid modeling, integrating process-based models with machine learning. Through open, reproducible research and compelling visualizations, I bridge the gap between cutting-edge research and societal impact.