Machine learning methods in Geophysics
The research of the working group "Machine Learning Methods in Geophysics" deals with the application of machine learning algorithms to various geophysical, mainly seismological, problems.
One of the biggest scientific challenges in the field of geophysics is the extraction of relevant information from the constantly growing amount of collected data. To tackle this problem, the "Machine Learning in Geophysics" working group analyzes geophysical data sets using probabilistic, data-driven approaches. On the one hand, this enables us to better understand the inherent structure of the data and, on the other, it helps us to evaluate large data sets quickly and objectively.
The data sets we deal with in the working group are mainly seismological signals. We automatically evaluate these, for example, with regard to the various signal sources that occur in them, such as earthquakes, landslides or even icequakes. In addition, we try to find correlations between different observations in order to estimate interesting parameters, such as the expected ground movement in the event of a certain excitation or the driving factors for the occurrence of an icequake.
If you would like to learn more about the research of the working group or about possible topics for Bachelor and Master projects, please contact Prof. Dr. Conny Hammer
Deep learning wavefield decomposition for diffraction imaging
Seismic and electromagnetic waves are not only reflected at large-scale horizons in the subsurface, but they are also scattered or diffracted at small-scale structures. These small-scale structures, such as faults, boulders or water intrusions in glaciers are crucial for the characterization of the subsurface. The separation of the often faint diffraced wavefield is therefore a key challenge in subsurface imaging. We utilize methods of machine learning for the separation of the diffracted wavefield in seismic and ground-penetrating-radar data.
Deep learning prediction of earthquake wavefields
The seismic wavefield that is excited by large earthquakes can be measured worldwide. A detailed knowledge of this wavefield is crucial for high-precision measurements that are carried out for instance in the field of particle physics. However, the regional distribution of seismological stations is often relatively sparse, particularly in regions with low seismic hazard such as Northern Germany. Based on synthetic earthquake waveforms, which can be computed for arbitrary points on the Earth's surface but rely on a relatively simple Earth model we utilize methods of machine learning to predict the seismic wavefield also at points, where no measurements are available.
Clustering of Large Distributed Acoustic Sensing Datasets
Distributed Acoustic Sensing (DAS) turns an optical fiber into a strainmeter by measuring backscattered light inside the fiber. It provides high-resolution seismic data with typical sampling frequencies in the order of 1 kHz and spatial sampling intervals of a few meters. The optical fiber can be several kilometers long and continuous measurements can be carried out over weeks, months or years. This can easily result in several terabytes of data per day, making it very challenging to obtain a good overview of what types of seismic signals the data contains.
Our aim is to cluster large DAS datasets in order to automatically detect and classify different types of seismic signals.
Identifying and Monitoring Different Types of Icequakes in Antarctica
A multitude of physical processes occurring within the ice of Antarctica give cause for the ice to vibrate. This study aims to identify the various seismic patterns generated within the ice and to relate them to potential processes. Furthermore, the objective is to monitor the occurrence of these processes over the last 20 years in order to gain insight into the influence of climate change on ice dynamics in Antarctica.