2024-07-12 –, While Loop (4.2)
Sindbad.jl currently serves as an internal modular framework for model-data integration at the Max Planck Institute for Biogeochemistry. It supports various methods for water-carbon coupling and vegetation dynamics. All definitions are compatible with automatic differentiation, enabling hybrid modeling approaches. Specifically, these approaches preserve the physics-based models while utilizing neural networks to optimize quantities of interest.
The application of automatic differentiation and deep learning approaches to tackle current challenges is now a widespread practice, and the biogeosciences community is no stranger to this trend. In this talk, we showcase the Sindbad.jl framework for modelling the ecosystem dynamics of vegetation, water, and carbon cycles. Our goal is to simulate targeted observations through physics-based models while employing a feedforward neural network to learn the spatial variability of their parameters. These models historically posed challenges due to their complex process representations, varied spatial scales, and parameterizations.
Our results demonstrate that the hybrid approach effectively predicts model parameters using a single neural network, compared to a site-level optimized set of parameters. This approach proves its capability to generate predictions consistent with in-situ parameter calibrations across various spatial locations, showcasing its versatility and reliability in modeling coupled systems.
This study illustrates that incorporating neural networks into a diverse collection of physics-based models holds significant promise. This integration has the potential to leverage the abundance of current Earth observations, enabling the application of these methods on a larger scale.
In addition to addressing scientific questions, we will also present technical challenges and the approaches taken. This includes dealing with mutation and automatic differentiation, scaling up forward runs with YAXArrays.jl, and ensuring documentation is user-friendly, or at least as user-friendly as possible.
Felix Cremer received his diploma in mathematics from the University of Leipzig in 2014. In 2016 he started his PhD study on time series analysis of hypertemporal Sentinel-1 radar data.
He is interested in the use of irregular time series tools on Synthetic Aperture Radar data to derive more robust information from these data sets.
He worked on the development of deforestation mapping algorithms and on flood mapping in the amazon using Sentinel-1 data.
He currently works at the Max-Planck-Institute for Biogeochemistry on the development of the JuliaDataCubes ecosystem in the scope of the NFDI4Earth project. The JuliaDataCubes organisation provides easy to use interfaces for the use of multi dimensional raster data
Lazaro Alonso is a Mexican physicist currently working at the Max Planck Institute for Biogeochemistry in the Model Data Integration Group. Interested in Hybrid Model-Based approaches to climate sciences as well as scientific visualization. Other interests of his are complex networks, graph neural networks and time series analysis.
He is a coauthor of the Julia Data Science book and main contributor to the gallery https://beautiful.makie.org/ and contributes as much as possible to open source in his spare time if any.
Scientist/Leader model-data integration group. PhD in Env. Sci. Eng.