← Back to Sessions

Global Source-Encoded Waveform Inversion: Preliminary Results

Description: 

We present the latest update on applying source encoding on a global scale. Our source-encoding technique enables us to compute Fréchet derivatives for all events with only a few forward and adjoint simulations and shows promising results for a hemispherical-scale study. In this study, we introduce four improvements. First, we start with the GLAD-M35 model, which is our latest global adjoint tomography model and provides a low-rank approximation of the full Hessian. This enables us to properly scale partial velocity Hessian terms as the preconditioner and determine the smoothing length based on the resolution length estimation. Second, we expand the dataset to a total of 9,382 events with corrected source mechanisms from 3D numerical simulations. The dataset is selected using a combination of FLEXWIN and new frequency-domain criteria, which restrict the difference between source-encoding and inversions using the full-frequency band. Third, by re-running forward simulations for all events every 10 iterations, we introduce a 'windowing' technique that removes unwanted parts of the time-domain observed data even when we only have a few data points in the frequency domain. Finally, we adopt new optimizers that better fit the varying nature of the source-encoded objective function. The optimizers we are testing include ADAM, stochastic conjugate gradient, and stochastic BFGS.

Session: 3D Wavefield Simulations: From Seismic Imaging to Ground Motion Modelling - III

Type: Oral

Date: 5/2/2024

Presentation Time: 02:15 PM (local time)

Presenting Author: Congyue

Student Presenter: Yes

Invited Presentation: Yes

Authors