Massively parallel time-domain FWI (Geoinv3D_fwk code)

We develop a massively-parallel time-domain FWI framework written in an object-oriented way to be able to tackle large-scale applications and plug easily new functions in the code. The code is developed by L. Combe at Geoazur. The figure below shows the architecture of the framework. The computations are managed by two MPI communicators to distribute the sources over the MPI process and to manage the domain decomposition of the subsurface model.

The modelling engine currently relies on finite difference method but can host alternative different numerical schemes on unstructured mesh thanks to a domain decomposition library. The optimization relies on the SEISCOPE toolbox developped by L. Métivier.

Ongoing work deals with the implementation of nonsmooth regularization.

Figure 1: Architecture of the time-domain FWI framework.

Figure 2 shows some results obtained on the Valhall oil field (Fig. 2). These results are compared with those obtained with FFWI.

Figure 2: Comparison between the results of visco-acoustic VTI FFWI (left) and acoustic VTI Geoinv3D_fwk (right). The preconditioning in depth behaves differently in the frequency and time domains although they rely on similar principles, the acquisition footprint is also different in the two sets of images. Accounting for attenuation in the left might have mitigated some artifacts in the gas cloud. We try to learn from the two implementations to better understand the sensitivity of FWI to several key functions.