Advances in Trans-dimensional Geophysical Inference
This research presents a series of novel Bayesian trans-dimensional methods for geophysical inversion. A first example illustrates how Bayesian prior information obtained from theory and numerical experiments can be used to better inform a difficult multi-modal inversion of dispersion information fr...
Main Author: | |
---|---|
Format: | Doctoral or Postdoctoral Thesis |
Language: | English |
Published: |
Canberra, ACT : The Australian National University
|
Subjects: | |
Online Access: | http://hdl.handle.net/1885/141087 https://doi.org/10.25911/5d5144b3620bd https://openresearch-repository.anu.edu.au/bitstream/1885/141087/4/Hawkins%20Thesis%202018.pdf.jpg |
Summary: | This research presents a series of novel Bayesian trans-dimensional methods for geophysical inversion. A first example illustrates how Bayesian prior information obtained from theory and numerical experiments can be used to better inform a difficult multi-modal inversion of dispersion information from empirical Greens functions obtained from ambient noise cross-correlation. This approach is an extension of existing partition modeling schemes. An entirely new class of trans-dimensional algorithm, called the trans-dimensional tree method is introduced. This new method is shown to be more efficient at coupling to a forward model, more efficient at convergence, and more adaptable to different dimensions and geometries than existing approaches. The efficiency and flexibility of the trans-dimensional tree method is demonstrated in two different examples: (1) airborne electromagnetic tomography (AEM) in a 2D transect inversion, and (2) a fully non-linear inversion of ambient noise tomography. In this latter example the resolution at depth has been significantly improved by inverting a contiguous band of frequencies jointly rather than as independent phase velocity maps, allowing new insights into crustal architecture beneath Iceland. In a first test case for even larger scale problems, an application of the trans-dimensional tree approach to large global data set is presented. A global database of nearly 5 million multi-model path average Rayleigh wave phase velocity observations has been used to construct global phase velocity maps. Results are comparable to existing published phase velocity maps, however, as the trans-dimensional approach adapts the resolution appropriate to the data, rather than imposing damping or smoothing constraints to stabilize the inversion, the recovered anomaly magnitudes are generally higher with low uncertainties. While further investigation is needed, this early test case shows that trans-dimensional sampling can be applied to global scale seismology problems and that previous analyses ... |
---|