Deep Earth Imaging – Future Science Platform

We use recorded seismograms at surface to image the subsurface structures across different scales. Seismic full waveform inversion is a state-of-art technique by using the complete recorded waveform information for inverting high-resolution subsurface velocity models. Wave-equation based full waveform inversion is a strong nonlinear problem. The traditional gradient-based linear approach heavily relies on a good starting model. A customized trans-dimensional Markov Chain Monte Carlo software is developed to allow us build pseudo-2D earth velocity model, without replying on the starting model, at the same time, providing uncertainty information. The algorithm is embarrassingly parallelized using MPI and OpenMP, thus is a perfect fit for HPC facilities. We use simulated data for simple and complicated model test. Field data test will be conducted this year.
Person

Principal investigator

Erdinc Saygin peng.guo@csiro.au
Magnifying glass

Area of science

Applied Science, Earth Sciences, Geosciences
CPU

Systems used

Magnus
Computer

Applications used

Competitive merit
Partner Institution: CSIRO| Project Code: pawsey0207

The Challenge

Seismic full waveform inversion is a state-of-art technique by using the complete recorded waveform information for inverting high-resolution subsurface velocity models. Wave-equation based full waveform inversion is a strong nonlinear problem. The traditional gradient-based linear approach heavily relies on a good starting model. Without a good starting model, the inversion may converge to a local minima

The Solution

In this study, we develop a nonlinear inversion approach. A customized trans-dimensional Markov Chain Monte Carlo software is developed to allow us build pseudo-2D earth velocity model, without replying on the starting model, at the same time, providing uncertainty information.

The Outcome

The algorithm is embarrassingly parallelized using MPI and OpenMP, thus is a perfect fit for HPC facilities, such as Magnus at Pawsey. The allocation allows us to run several tests using simulated dataset. We need more allocations in future to run more complicated model test and for field data test.

List of Publications

Bayesian Transdimensional Seismic Full Waveform Inversion with a Dipping Layer Parameterization, Visser, Guo and Saygin, submitted to Geophysics