Numerical Assessment of a Multi-Step One-Shot Approach for Frequency-Domain Acoustic FWI

Full waveform inversion (FWI) aims to reconstruct spatially parameters of a wave propagation model by minimizing a misfit between simulated and measured wavefields. The process generally involves two main components: a numerical solution stage for the forward and adjoint wave problems (to estimate the gradient of the misfit), and an outer optimization algorithm. In large-scale scenarios, the wave problems are solved using iterative domain decomposition techniques, and local gradient-based methods are used as optimization algorithms. The one-shot methodology modifies the traditional approach by intertwining the solution and optimization steps. Instead of fully resolving the forward and adjoint systems at each optimization iteration, it uses a limited number of solver iterations, yielding an approximate gradient. This presentation explores the viability and effectiveness of such a one-shot strategy for frequency-domain FWI when an Optimized Restricted Additive Schwarz (ORAS) method is combined with a Gauss-Newton optimization strategy. The analysis focuses on computational performance, considering the number of solver iterations, gradient updates, and subdomain partitions. This is joint work with Boris Martin, and Christophe Geuzaine.