Accelerated large-scale inversion with message passing

Abstract

To meet current-day challenges, exploration seismology increasingly relies on more and more sophisticated algorithms that require multiple paths through all data. This requirement leads to problems because the size of seismic data volumes is increasing exponentially, exposing bottlenecks in IO and computational capability. To overcome these bottlenecks, we follow recent trends in machine learning and compressive sensing by proposing a sparsity-promoting inversion technique that works on small randomized subsets of data only. We boost the performance of this algorithm significantly by modifying a state-of-the-art ℓ1-norm solver to benefit from message passing, which breaks the build up of correlations between model iterates and the randomized linear forward model. We demonstrate the performance of this algorithm on a toy sparse-recovery problem and on a realistic reverse-time-migration example with random source encoding. The improvements in speed, memory use, and output quality are truly remarkable.