Earthquake source inversion with dense networks

Surendra Nadh Somala, Jean-Paul Ampuero, & Nadia Lapusta

Published December 2012, SCEC Contribution #1882

Inversions of earthquake source slip from the recorded ground motions typically impose a number of restrictions on the source parameterization, which are needed to stabilize the inverse problem with sparse data. Such restrictions may include smoothing, causality considerations, predetermined shapes of the local source-time function, and constant rupture speed. The goal of our work is to understand whether the inversion results could be substantially improved by the availability of much denser sensor networks than currently available. The best regional networks have sensor spacing in the tens of kilometers range, much larger than the wavelengths relevant to key aspects of earthquake physics. Novel approaches to providing orders-of-magnitude denser sensing include low-cost sensors (Community Seismic Network) and space-based optical imaging (Geostationary Optical Seismometer). However, in both cases, the density of sensors comes at the expense of accuracy. Inversions that involve large number of sensors are intractable with the current source inversion codes. Hence we are developing a new approach that can handle thousands of sensors. It employs iterative conjugate gradient optimization based on an adjoint method and involves iterative time-reversed 3D wave propagation simulations using the spectral element method (SPECFEM3D). To test the developed method, and to investigate the effect of sensor density and quality on the inversion results, we have been considering kinematic and dynamic synthetic sources of several types: one or more Haskell pulses with various widths and spacings; scenarios with local rupture propagation in the opposite direction (as observed during the 2010 El Mayor-Cucapah earthquake); dynamic crack-like rupture, both subshear and supershear; and rupture that mimics supershear propagation by jumping along the fault. In each case, we produce the data by a forward SPECFEM3D calculation, choose the desired density of stations, filter the data to 1 Hz (since the bulk properties are not known at higher frequencies), add noise of the desired level, and then apply our inversion approach. The results indicate that dense networks (e.g., 1-km spacing) produce sharper images of the considered sources than sparse networks (e.g., 10-20 km spacing), with better amplitude recovery and better resolution with depth. This is true even when noiseless sparse networks are compared with noisy dense networks, provided that the standard deviations of noise do not exceed ~1% of the maximum earthquake source amplitude (e.g., 1 cm/s noise for 1 m/s Haskell source). Substantial qualitative improvements arise when features of relatively narrow spatial extent are included in the source, in which case the dense networks can reproduce the features whereas the sparse networks cannot. We will report on our current efforts to mathematically quantify the differences between the inversions of sparse and dense data and to incorporate the effect of errors in the bulk velocity model.

Somala, S., Ampuero, J., & Lapusta, N. (2012, 12). Earthquake source inversion with dense networks. Oral Presentation at AGU Fall Meeting 2012.