GITEC: A Generalized Inversion Technique Benchmark

Abstract

ABSTRACT Generalized inversion techniques (GITs) have become popular for determining seismological parameters (e.g., source, attenuation, and site response), particularly in low-to-moderate seismicity regions. Indeed, GITs can potentially provide reliable site-response estimates when a minimum number of recordings is available, as well as valuable information about source parameters and regional attenuation characteristics. Significant advances have been made on GITs in which different approaches and hypotheses were investigated, such as the application of “nonparametric” and “parametric” inversion schemes. In this context, several scientific questions have arisen that depend on the final scope of the GITs: What is the optimal inversion strategy for a given dataset configuration? What is the impact of the different choices, assumptions, and implementations on the reliability of the results? Is it possible to quantify the associated epistemic uncertainties? Here, we have considered and compared the different approaches of GITs to improve the understanding of each for use in different applications. A methodological benchmark that includes different GIT methods and dataset configurations is set up to fulfill the objective, using a simple synthetic dataset, a French regional sparse dataset, and an Italian national dense dataset. The benchmark is developed in two phases: (1) phase I: a free phase with no common constraints; and (2) phase II: a constrained phase with unified reference conditions. Despite unifying the reference conditions in the different inversions, the variability was not reduced. Discrepancies are observed between different terms of GITs. Site responses appear to be the most robust estimates, compared to source and attenuation terms. The way that stress drops of earthquakes and quality factors for crustal attenuation are parameterized appears to lead to significant variability between different approaches. Finally, uncertainties are addressed by quantification of the inter-method variability for the different terms and parameters.