Evolutionary full waveform inversion

Common full-waveform inversion (FWI) workflows rely on a static snapshot of the available earthquake data, and it is not always obvious how information from new events should be incorporated. In this contribution we introduce several methodological advancements which allow for the continuous inclusion of new earthquake data, resulting in a dataset and model which "evolve" over time. We demonstrate the utility of such approaches in an ongoing evolutionary FWI of the African crust and mantle.

Our contributions can be broken down into the following workflow. First, data-centers are queried automatically to expand the current data set, after which recordings are processed and windows are selected. The entire workflow from data acquisition to model update is fully automated. To accommodate the dynamic nature of an evolving dataset we compute model updates with one of the following two strategies. (1) Batches of events with a good spatial distribution are chosen from a complete dataset in a randomised manner to compute the model update. The batch size is adapted to the amount of additional information each individual gradient provides to the sum of gradients present in the batch. (2) When the model has converged to a minimum for all previously available data, the gradient due to a new event will dominate and requires the computation of the gradient of the new event only. Previously calculated gradients can still be used to approximate the inverse Hessian in Quasi-Newton optimization methods.

These developments are highlighted in an application of evolutionary full-waveform inversion of the African continent, which is melded together within the framework of the Collaborative Seismic Earth Model (CSEM) project. Recordings from the AfricaArray and NARS seismic networks are complemented with data from global networks.