Ambient seismic noise imaging using trans-dimensional and machine learning techniques with application to Borneo and Iceland

The exploitation of surface wave dispersion measurements from ambient seismic noise data for the purposes of Earth imaging has become increasingly common in recent years and has led to many discoveries that have changed or re-enforced our understanding of Earth structure and processes. At the same time, computational advances have led to the use of sophisticated Bayesian trans-dimensional inversion schemes, which allow for variable resolution and can estimate posterior uncertainty, though at increased computational cost compared to traditional linearised methods. In the first part of this thesis, I leverage ambient noise data and trans-dimensional methods to investigate the crustal structure of northern Borneo, a post-subduction setting in which sequential but opposed subduction systems recently terminated. In this study I used a two-step trans-dimensional method that also included a neural network inversion scheme in the second step. The neural network improved the final model by producing a smoother 3D structure with fewer artefacts without sacrificing the fit to the data. The final model reveals underthrust continental material beneath the Crocker Range and a low velocity anomaly at depth in a region of crustal thinning. In the latter case, this was interpreted as a region of thermal upwelling associated with failed rifting in northern Borneo that was initiated by opening of the Sulu Sea to the north due to roll-back of the Celebes Sea slab. In the second part of the thesis, I implement two-step trans-dimensional inversion of ambient noise surface wave data extracted from a dense (~1 km station spacing) network at Askja volcano in Iceland. Here, I computed velocity models for both Vsh (from Love waves) and Vsv (from Rayleigh waves), along with associated uncertainties, which allowed radial anisotropy to be measured. This model reveals a pronounced low Vs anomaly at 1-3 km depth beneath the caldera, which I interpret to be the signature of a shallow magma chamber. A positive radial anisotropy anomaly in the same location suggests that the magma chamber is made up of horizontal sill-like structures with an aspect ratio (vertical/horizontal) of between 0.2-0.8. In the final part of the thesis, I develop a neural network emulator for the one step inversion forward problem in order to greatly improve computational efficiency, thus making time-consuming inversion methodologies, such as Bayesian McMC inversions, more tractable. I build and train this emulator from a set of synthetic data sampled from a pre-defined prior. The errors in the emulator are low and the computation of a single forward problem is ~100 times faster than the deterministic approach. I employ the automatic gradient calculations inherent in neural networks, which is an advantage of this method, and perform an example synthetic inversion for 3-D structure, which is able to recover the primary features of the input model.

Studentship funded jointly by the Engineering and Physical Sciences Research Council (EPSCR) and CGG (G106858)