Seismic rotational data: Acquisition, processing and applications. SEP-167
Table of contents
- Chapter 1: Introduction
- Chapter 2: Seismic rotations from induction-coil magnetometers
- Chapter 3: Wave-mode separation with translations and rotations
- Chapter 4: Automatic wave mode identification using machine learning
Rotations of the ground as a result of the propagation of seismic energy are one of the least-studied phenomena in seismology. This is curious, since unlike other, more complicated geophysical theories, rotations are an observable that we can measure on the Earth's surface. Furthermore, the concept of medium rotation is inherent to elastic continuum theory, and has been known to exist for as long as seismic measurements have been made. In many fields of technology (the aerospace and gaming industry, for example), it is well-understood that to measure motion of a finite body it is necessary to measure both 3D translations and 3D rotations.
However, the industry-standard measurements in exploration and in earthquake seismology have always been of translations on land, pressure in water, or both translations and pressure on the seabed. There is currently no technological solution which would enable the seismological community to record rotational data on a scale similar to the recording of translations.
Part of the reasons for this are circumstantial: the instruments that were more readily available at the time that basic seismological research was taking place measured translations, not rotations. These instruments were improved over the years, leading to the development of the robust geophones and accelerometers common today, with their high sensitivity and wide dynamic range. In marine seismic acquisition, a sensor that measures pressure is the obvious choice, and indeed hydrophones are the mainstay of marine seismic data acquisition. Furthermore, the design of an instrument which would record only rotations and yet would be insensitive to translations is not obvious.
The availability of a particular type of observable data inspired research to acquire and process these data for imaging of the Earth's subsurface (exploration seismology) and for deriving earthquake source mechanisms (earthquake seismology). Geophysical research (and the resulting extraction of Earth's mineral resources) has been very successful using existing translational and pressure data. So successful, in fact, that the mere idea that the data are incomplete appears either prepostorous or irrelevant upon initial inspection.
The elastodynamic seismic wavefield contains more than just three components of translations and one of pressure. However, until very recently the seismic industry has not asked itself the very simple question: “what observables out of the elastodynamic seismic wavefield would we ideally like to record ?”
My purpose in this dissertation is to show how to include rotational data in the seismic acquisition and processing workflow. I use the term “6C data” to refer to data comprising three components of translations and three components of rotation, and “7C data” if hydrophone data are also included.
First, I cover some of the current rotational-acquisition methods, discussing their benefits and shortcomings. I then propose using existing induction-coil magnetometer technology to record rotations, and show a field data experiment where this possibility is validated.
The combination of rotational motion records together with current translational motion records can aid in the processing of seismic data by providing extra information about the seismic wave modes being recorded. I show one application where particular wave modes are selected and separated from a multicomponent field dataset that has both translational and rotational components.
The use of machine learning algorithms in exploration seismology is not new, and is currently seeing a resurgance due to the rapid increase in data volumes and complexity. I show how to extract feature vectors from a multicomponent dataset comprising both translational and rotational components, and use them as a training set for a machine learning algorithm. This algorithm can then identify particular wave modes based on this training. The possibility of using a machine learning algorithm to identify and separate wave modes from large seismic data volumes is compelling, as it has the potential for saving on time-consuming manual processing steps.
Reproducibility and source codes
This thesis has been tested for reproducibility. The source codes and makefiles to generate the resutls shown in chapter 3 of the thesis can be reproduced by cloning the git repository. You will then need to download and install Docker. Follow the instructions in the README.md file in the git repository in order to generate the runtime environment, compile and run the code, and generate the figures seen in chapter 3.