Humboldt-Universität zu Berlin - Mathematisch-Naturwissenschaftliche Fakultät - Institut für Informatik

Dissertation: Jannes Münchmeyer

Am Donnerstag, 10.03.2022 wird Jannes Münchmeyer sein Dissertationsthema in einem Vortrag vorstellen. Der Titel der Arbeit ist:

  "Implications for rupture predictability and early warning"

Der Vortrag findet am Donnerstag, 10.3.22, zwischen 09:00-11:00 Uhr statt – voraussichtlich in hybrider Form.

 

Abstract

Machine learning for fast and accurate assessment of earthquake source parameters: Implications for rupture predictability and early warning

Earthquakes are among the largest and most destructive natural hazards known to the humankind. While records of earthquakes date back millennia, and systematic studies of earthquakes have been conducted for over a century, many questions about their nature remain open. One particularly interesting question is termed rupture predictability: to what extent is it possible to foresee the final size of an earthquake while it is still ongoing? This question is integral to earthquake early warning systems trying to provide information about ongoing earthquakes to places where shaking has not yet arrived, allowing for last moment preparatory action. Still, research on this question so far has reached contradictory conclusions.

In recent years, the advent of big data and big data analysis techniques opened up novel research opportunities for gaining insights into rupture predictability. The amount of data available for earthquake research has grown exponentially during the last decades, as for many other scientific domains, reaching now tera- to petabyte scale, with future growth to be expected. This wealth of data, while making many approaches based on manual inspection infeasible, now allow for data driven analysis and complex models with high numbers of parameters. As special class of theses models are machine learning, in particular deep learning, methods. Deep learning has gained overwhelming interest across domains in the last decade, driven by new developments in the field. It already lead to considerable improvements upon previous methods for many seismic analysis tasks. Nonetheless, the application of deep learning methods to seismological observables is still in its infancy.

In this thesis, we develop machine learning methods for the study of rupture predictability and the closely related task of earthquake early warning. We first study the calibration of a high-confidence magnitude scale in a post hoc scenario. For this, we develop a hybrid approach, based on mathematical optimisation and machine learning. Subsequently, we focus on real time estimation models, based on deep learning. We develop the transformer earthquake alerting model (TEAM), a method for earthquake early warning, estimating ground motion parameters directly from seismic waveforms. TEAM outperforms traditional early warning methods in terms of warning times and the relation between true, false and missed warnings. Based on TEAM, we develop TEAM-LM, a model for real-time magnitude and location estimation. We use TEAM-LM to study the advantages and shortcomings of deep learning methods for the assessment of earthquake source parameters in comparison to classical approaches. Our analysis shows that such methods outperform classical approaches consistently when provided with sufficient training data, but exhibit systematic mispredictions for data-sparse scenarios. As a particular relevant case of data sparsity, we identify large magnitude events. In a last step, we use TEAM-LM and the insights gained through its analysis to study rupture predictability. For this, we collate a dataset of teleseismic P wave arrivals, encompassing events and stations worldwide. We complement this analysis with results obtained from a deep learning model based on moment rate functions. Our analysis shows that earthquake ruptures are not predictable early on, but only once their peak moment release has been reached, after approximately half of their duration. Even then, potential further asperities can not be foreseen. While this thesis finds no rupture predictability, the methods developed within this work demonstrate how deep learning methods make high quality real time assessment of earthquakes practically feasible. We hope that these results will allow to improve future earthquake early warning systems, and thereby help to reduce the harm caused by earthquakes.