• Important information on external master thesis projects!

• Allg. Hinweise zur Form der Abschlussarbeiten am MIM

Bei Interesse an Masterarbeiten im Lehrstuhl für Theoretische Meteorologie bitte bei Prof. G. Craig, Prof. T. Birner, oder Dr. C. Keil nachfragen. Nachfolgend eine Auswahl derzeit offener Themen:

The European Centre for medium-range weather forecasting (ECMWF) has accidentally run a one year-long butterfly experiment, with their deterministic and control forecast initially differing only by tiny truncation errors. This data set, together with the regular ensemble forecasts, provides a great opportunity to study uncertainty growth in weather prediction and to address related questions on forecast improvement potential and forecast busts. The thesis requires good python coding skills and interest in data analysis and visualization.

For more detailed information please contact Tobias Selz or George Craig.

Piece-wise potential vorticity (PV) inversion is a method to understand spatial interactions in the atmosphere. It has been used to investigate the reasons for uncertainty growth or extreme weather. However, running the numerical inversion algorithm on forecast data is expensive and often leads to convergence issues. The goal of the thesis is to investigate, if AI-based methods could be used for the PV inversion instead. Therefore, a training data set has to be created first, applying the conventional method. After that, various neuronal network designs should be developed, trained and tested. The thesis required excellent python coding skills and working independently with neural networks.

For more detailed information please contact Tobias Selz or George Craig.

Short-range precipitation forecasts often exhibit overconfidence in ensemble forecasts, leading to discrepancies with observed precipitation extending beyond the predicted area. Incorporating model uncertainty schemes may mitigate this issue by introducing greater spatial variability. Spatial variability can be assessed using spatial skill scores such as the Fractions Skill Score (FSS) and Structure-Amplitude-Location (SAL) score, applied to different combinations of ensemble members. Since these scores involve squared differences between ensemble members, they can be reformulated in a form of variance. Analyzing how this variance is modulated by including model uncertainty schemes can provide insights for improving ensemble forecasts. This study examines how the spatial variability of precipitation forecasts is modified by a physically based stochastic perturbation scheme, with a particular focus on hourly precipitation during the summer over Germany. The variance budget framework proposed by Matsunobu et al. (2025) will be extended to incorporate FSS- and SAL-based spread metrics.

For more detailed information please contact Takumi Matsunobu or Christian Keil.

For more detailed information please contact Tobias Selz or George Craig.

The stratospheric circulation is projected to change in response to a warming climate. A changing circulation can impact the transport of trace gases (ozone, water vapor and ozone destroying substances) in the middle and upper atmosphere. The project will focus on developing a better understanding of the dynamics-transport coupling in the stratosphere.

Changes in the large-scale stratospheric dynamics and transport of trace gases will be investigated using a combination of theory, observations and numerical modeling. Key topics of investigation will be the subseasonal to climatological impact of stratospheric dynamical processes on the global distribution of tracers in the stratosphere. A hierarchy of climate models will be used.

For more detailed information please contact Thomas Birner.

Although convective clouds are much smaller than large-scale weather systems, they can be a major source of error in weather forecasts because errors grow rapidly. Studying this error growth requires large ensembles of forecasts to accurately represent the many ways that the weather situation can develop. The aim of this project is to develop a simple numerical model that can represent the error growth processes, but is inexpensive to run in large ensembles. Different model formulations will be designed, programmed in Python and evaluated using advanced verification measures.

For more detailed information please contact George Craig.

Topics broadly in the area of stratosphere-troposphere and climate dynamics are available upon request. Recent research topics in our group include: variability and long-term trends in the width of the tropical belt, processes that govern the temperature structure of the tropical tropopause layer, the dynamics of sudden stratospheric warmings and their coupling to the troposphere, transport processes in the upper troposphere / lower stratosphere. Interested candidates are asked to look through research topics on our group's website, in particular our recent publications: https://www.meteo.physik.uni-muenchen.de/~Thomas.Birner/pubs.html.

For more detailed information please contact Thomas Birner.

The assimilation of cloud-related observations is challenging as traditional error metrics (e.g. RMS error) are not really suitable for the evaluation of intermittent fields as clouds or precipitation (double penalty problem). To overcome this deficiency, various feature-based scores have been developed for the verification of forecast precipitation fields, but these scores have not yet been applied in the context of data assimilation.

The goal of this thesis ist to test the use of feature-based metrics for the assimilation of cloud-affected satellite observations in the emsemble data assimilation system KENDA for the regional weather forecast model COSMO-DE. Different approaches shall be tested in an idealized setup of KENDA that is used by several people in the HErZ data assimilation group.

Contact: Leonhard Scheck

The storage space requirements for output of numerical weather and climate prediction is growing faster than the cost of storage space is decreasing. Model resolution is continuously increased to overcome issues with parameterized physical processes like convection. At the same time the ensemble size (e.g. the number of model runs performed to create one forecast) is also increased to improve the assessment of uncertainty within the forecast. Both in combination results in really big data sets that are not only difficult to process but also very expensive to store.

One way to reduce the amount of output is to rely more on online diagnostics and not to save large fractions of the output. While this approach appears promising in operational setups of weather services, it is only a partial solution in research. Visualization of arbitrary aspects of a model run would no longer be possible and experiments would have to be carried out again if changes are made to the online diagnostic.

Another way is to store model output with reduced precision. File formats currently in use support only lossless compression or if lossy compression is possible only spatial correlation between neighboring data points is used (e.g., JPEG compression). The temporal correlation is ignored. In contrast, video compression algorithms are essentially based on the temporal correlation between successive time steps. Without reducing the quality, this results in a compression ratio that is by one order of magnitude higher than that of individual images. Central ideas of video compression should be directly transferable to the compression of model output. Examples are differential coding (only differences between time steps are stored) and motion compensation (for unchanged but moved parts of an image only a displacement vector is stored). On the other hand, assumptions about the perception by the human eye are not applicable (e.g., changes in brightness and color are not equally important).

The following questions should be addressed in this thesis:
Which compression algorithms are best suited for meteorological model output? Candidates are video compression and general purpose algorithms. The plan is not to develop new algorithms, but to asses existing ones.
What are the characteristics of errors produced by the analyzed algorithms?
Which degree of compression is acceptable for a set of different meteorological applications?

Contact: Robert Redl

Further themes are possible, please talk to Prof. G. Craig, Prof. T. Birner, or Dr. C. Keil.