Machine Learning and Modeling Techniques for Regional Seismology – An Application to Southern Europe

Note: This project is designed for up to two students. 

Project description

Motivation (background):
Volcanic eruptions and earthquakes are poorly understood hazards that pose significant direct and indirect (for example through triggering of landslides and tsunamis) risks to humans, the environment and the economy. A better understanding of both volcanic eruptions and earthquakes requires more accurate seismic velocity models. Moreover, the tectonic regions in which these hazards typically occur (for example in southern Europe), often have complicated plate boundaries, including deformed and torn subduction zones, which are believed to strongly influence mantle flow and plate motions. Improved seismic velocity models are therefore key in getting a better understanding of hazards, geodynamics and how these are related.

Hypothesis (scientific problem):
The seismic data used to obtain these velocity models most often are travel times.
In order to improve the resolution of the velocity models it is important to get more and more accurate travel time picks as well as amplitudes and then to use these in modeling and inversion algorithms that are both better and faster than available algorithms. The former (picking of travel times and amplitudes) is the focus of the first master project and the latter (developing improved modeling and inversion algorithms) is the main topic of the second master project.

Test (work):
Travel times are routinely picked by experienced analysts. However, there is a need for automated picking which is motivated by a number of factors: the large increase of data (as provided by AdriaArray), the complexity and frequency dependence of the picks and the need to get as many reliable picks as possible in the presence of noise. A first step in this direction has been the development of automated picking routines. A second step, and the focus of this first thesis, is the use of Machine Learning algorithms to further improve the picks. This requires, among other things, the development and testing of algorithms, such as convolutional neural networks. In principle these algorithms can be used to pick any type of phase, but special attention will be given to important crustal phases such as Pg and Sg and refracted phases such as Pn and Sn. 

The picked phases, and their related waveforms, contain a wealth of information on Earth structure as well as earthquake location and mechanism. In order to process these data systematically, accurate modeling and inversion techniques are required. Currently, this is typically done using ray tracing through 1D isotropic velocity models. In this thesis the focus will be on developing and testing more accurate ray tracers (for example, ray tracers that take anisotropy and attenuation into account) as well as related modeling algorithms, with special attention to the phases mentioned above. These will be used to create and invert synthetic data and, if there is time, to invert real data for a particular region of interest.

Proposed course plan during the master's degree (60 ECTS):
GEOV274 (10 sp) or AG335 (10 sp, geophysics field course in UNIS, Svalbard)
GEOV277 (10 sp)
GEO-DEEP course on mantle dynamics at UiO (5sp)
Geophysics field course (5sp)
GEOV302 (10sp) or GEOV276 (10sp)
GEOV359 (10sp)
INF264 (10sp)

Prerequisites
Bachelor in geophysics

Work
Both these projects happen in the context of AdriaArray, a large multi-annual geophysical project that started in 2020 and in which the university of Bergen, as well as the universities of Kiel and Cambridge are involved. As part of the AdriaArray project hundreds of seismometers are being installed in a region that covers the Alps and the Mediterranean (roughly from southeastern France to Crete and from Sicily to the Black Sea). This will significantly increase the amount of data in this area. Another part of the AdriaArray project consists of further developing and testing various modeling and inversion algorithms. The two master projects are therefore well integrated into the AdriaArray project.