A unified 2D/3D software environment for large scale time-harmonic full waveform inversion

Full-Waveform Inversion is a costly and complex procedure for realistically sized 3D seismic data. The performance-critical nature of this problem often results in software environments that are written entirely in low-level languages, making them hard to understand, maintain, improve, and extend. The unreadability of such codes can stymie research developments, where the translation from higher level mathematical ideas to high performance codes can be lost, inhibiting the uptake of new ideas in production-level codebases. We propose a new software organization paradigm for Full-Waveform Inversion and other PDE-constrained optimization problems that is flexible, efficient, scalable, and demonstrably correct. We decompose the various structural components of FWI in to its constituent components, from parallel computation to assembling objective functions and gradients to lower level matrix-vector multiplications. This decomposition allows us to create a framework where individual components can be easily swapped out to suit a particular user's existing software environment. The ease of applying high-level algorithms to the FWI problem allows us to easily implement stochastic FWI and demonstrate its effectiveness on a large scale 3D problem.

https://slim.gatech.edu/Publications/Public/Conferences/SEG/2016/dasilva...