Scaling full seismic waveform inversions

Krischer, Lion (2017): Scaling full seismic waveform inversions. Dissertation, LMU München: Fakultät für Geowissenschaften

Abstract

The main goal of this research study is to scale full seismic waveform inversions using the adjoint-state method to the data volumes that are nowadays available in seismology. Practical issues hinder the routine application of this, to a certain extent theoretically well understood, method. To a large part this comes down to outdated or flat out missing tools and ways to automate the highly iterative procedure in a reliable way. This thesis tackles these issues in three successive stages. It first introduces a modern and properly designed data processing framework sitting at the very core of all the consecutive developments. The ObsPy toolkit is a Python library providing a bridge for seismology into the scientific Python ecosystem and bestowing seismologists with effortless I/O and a powerful signal processing library, amongst other things. The following chapter deals with a framework designed to handle the specific data management and organization issues arising in full seismic waveform inversions, the Large-scale Seismic Inversion Framework. It has been created to orchestrate the various pieces of data accruing in the course of an iterative waveform inversion. Then, the Adaptable Seismic Data Format, a new, self-describing, and scalable data format for seismology is introduced along with the rationale why it is needed for full waveform inversions in particular and seismology in general. Finally, these developments are put into service to construct a novel full seismic waveform inversion model for elastic subsurface structure beneath the North American continent and the Northern Atlantic well into Europe. The spectral element method is used for the forward and adjoint simulations coupled with windowed time-frequency phase misfit measurements. Later iterations use 72 events, all happening after the USArray project has commenced, resulting in approximately 150`000 three components recordings that are inverted for. 20 L-BFGS iterations yield a model that can produce complete seismograms at a period range between 30 and 120 seconds while comparing favorably to observed data.

• Dokumententyp: Dissertationen (Dissertation, LMU München)
• Keywords: Seismology, Full Waveform Inversion, Optimisation, Wave Propagation, Adjoint-State Method, Data Formats, Big Data, Python, Data Science, Workflows
• Themengebiete: 500 Naturwissenschaften und Mathematik ; 500 Naturwissenschaften und Mathematik > 550 Geowissenschaften
• Fakultäten: Fakultät für Geowissenschaften
• Sprache der Hochschulschrift: Englisch
• Datum der mündlichen Prüfung: 19. Juli 2017
• 1. Bericht­erstatter:in: Igel, Heiner