High-resolution observations from instruments on geostationary satellites provide a wealth of information about convective activity and are therefore seen as a important type of observation for convective scale data assimilation (DA). In particular the visible and near-infrared channels provide information on the cloud distribution, cloud microphysical properties and cloud structure with high temporal and spatial resolution. However, in operational DA systems currently only clear sky thermal infrared and microwave radiance observations are used, which mainly provide temperature and humidity information. Sufficiently fast and accurate forward operators for visible and near-infrared radiances are not yet available, because multiple scattering makes radiative transfer at solar wavelengths complicated and computationally expensive.
Only recently a we developed MFASIS, a loop-up table based 1D radiative transfer method that is orders of magnitude faster than conventional radiative transfer solvers for the visible spectrum and similarly accurate. A preliminary version of a forward operator based on this method, which simulates synthetic MSG-SEVIRI images from COSMO-DE model output, has been completed and implemented in the pre-operational km-scale Ensemble Data Assimilation (KENDA) system of DWD. To improve the consistency and accuracy of the operator we included computationally cheap methods to account for the overlap of subgrid clouds and the most important 3D radiative transfer effects.
MFASIS has also been used to evaluate large domain large eddy simulations carried out using the ICON model in the framework of the HD(CP)2 project. Synthetic 0.6 and 0.8 micron MODIS images with 250m resolution were computed from 3D model states with a grid resolution of about 150m. The high resolution allowed us to determine the effective grid resolution from cloud size distributions.
Tropical cyclones, singular Vectors, adaptive grid refinement in the framework of the METSTRÖM project, high performance computing, visualization and computational steering at LRZ, core collapse supernovae and extragalactic jets at the MPI for Astrophysics.
Fornado = a FORtran Neural network inference code including an ADjOint version. Fornado can be used as a replacement for Tensorflows predict subroutine and is optimized for small networks (less than 100 nodes per layer).
CosmoState is a grib_api-based Python module to read GRIB1 and GRIB2 files generated by COSMO (and possibly other NWP models).
Interactive IPython notebook illustrating the properties of simple numerical schemes for the linear advection equation and the non-linear Burgers equation (there is also a non-interactive HTML version).
See also ORCID, ResearcherID/publons, ResearchGate or Google Scholar.