The motion of interfaces is a crucial part of the modeling and simulation in applications like two phase flow, crystallization and phase separation, or tumor growth. Within the Copernicus program, radar and optical satellite imagery are available with short revisit periods, allowing for high quality observation of moving fronts of wildfires, oil spills and floods, or calving fronts of glaciers and ice sheets. Front tracking in satellite data therefore is a key ingredient for various applications in Earth Observation and an active research topic in various Helmholtz centers. Apart from questions concerning image analysis, the geometric information contained in the moving fronts is essential for the mathematical models, which are the basis for the simulation and forecasting of complex physical phenomena in the Earth system.

The main objective of the TerraByte-DNN2Sim project is to develop novel imaging algorithms and a data processing pipeline for the automated extraction of moving fronts and further geometric information from satellite data to make this information available for applications in various research fields and all Helmholtz centers using satellite data, e.g. DLR, FZJ, UFZ, AWI. We use this extracted data in simulations to advance the modelling of front evolution that is important in many research areas. We apply these methods to models of glacier calving fronts of the antarctic ice sheet. Computation will be carried out on the new TerraByte cluster that provides efficient access to geoscientific data like satellite images as well as the computational resources necessary for simulation.

Given the time series of front positions extracted from image data, our goal is to compute the spacially and temporally resolved speed of the front. The front can be mathematically represented by the zero levelset of an implicit function. The evolution of the front is then described by a partial differential equation, which includes the speed of the front as a parameter. We will implement a software library to solve the inverse problem of computing the speed of the front from known front positions, taking care of the necessary regularization. From there it is possible to compute the position of the front at every time and forecast it into the future. The library should be efficiently parallelized and scale to thousands of CPUs. Capturing the moving front requires a dynamically adapted grid which poses a challenge to the parallelization. The library should support diverse use cases and allow easy integration into or coupling with other simulation codes, e.g., using the coupling framework preCICE.

Computations in the project will be performed on TerraByte, a new High Performance Computing cluster for earth observation and other geoscience workloads established by DLR and hosted at the Leibnitz Supercomputing Centre (LRZ) of the Bavarian Academy of Sciences and Humanities. The first stage of TerraByte consists of 61 compute nodes with 80 CPU cores each, 15 GPU nodes with 4 Nvidia GPUs, and 36 petabytes of data storage, including fast access to satellite images and other geoscientific data. A further significant extension of compute resources is planned for 2023. TerraByte therefore provides both the efficient data access for image analysis as well as the compute power required for simulation.