Abstract Elastic full-waveform inversion (FWI) has emerged as the standard methodology for structural and quantitative subsurface imaging. One of the most advanced implementations incorporates TTI anisotropy and attenuation. Its computational demands, however, create economic and timeline constraints that threaten commercial viability. These constraints were addressed through three approaches: high-performance computing infrastructure, optimized algorithms, and streamlined workflows that eliminate compute idle time. The computationally intensive wave propagation component, which dominates FWI runtime, was efficiently handled through graphics processing unit (GPU)-accelerated nodes using the latest NVIDIA GPU architectures. However, FWI as a workflow was not only a compute problem but also a data movement problem. As GPUs became faster, they exposed bottlenecks in other parts of the workflow. To prevent this, data movement among storage, memory, and compute units was optimized to reduce I/O bottlenecks. This was mainly done using diverse fit-for-purpose nodes and asynchronous processing. Our approach demonstrates how careful consideration of infrastructure characteristics, combined with algorithmic improvements, enables significant runtime reductions while maintaining computational accuracy and cost-effectiveness.
Datta et al. (Thu,) studied this question.