2025-10-02 –, Coffee room
Language: English
In this presentation, we showcase Parameter-Learning in soil carbon modeling, a scientific machine learning paradigm where a process-based model is parameterized through the use of a neural network. This enables us to learn latent soil properties from soil carbon data, leveraging the Julia’s ML and auto-differentiation ecosystems. The resulting hybrid model is more explainable through the latent variables, and more robust as it is constrained to a process-based model.
Soils are a crucial component of the global carbon cycle, and store more carbon than the atmosphere and vegetation combined. Therefore learning latent soil properties (such as microbial carbon use efficiency or sorption capacity) is important for understanding the fate of soil carbon under climate change.
However, ground soil sampling data is sparse, and only soil carbon data is available with a sufficiently comprehensive global distribution. Latent soil properties can nevertheless be inferred by parameterizing a process-based model that predicts soil carbon, where the parameters of the process-based model are our latent variables of interest. Previously, parameterizing these models for global predictions was a two-step process, first a data assimilation process, followed by training the ML model on the results of the data assimilation. Thanks to the recent progress of auto-differentiation frameworks, a one-step end-to-end approach is now possible; a neural network predicts parameters to be fed to a process-based model, and its weights are optimized through auto-differentiation to give fast, reliable, scalable results.
Contributors: Bernhard Ahrens, Leo Rossdeutscher, Marco Paina, Thomas Wutlzer, Marion Schrumpf, Markus Reichstein
First year PhD Researcher at the Max Planck Institute for Biogeochemistry with a background in theoretical physics. Julia user since 2017.
Research Summary: Parameterizing physical and other models using neural networks, applications to soil science.