Breaking the bottlenecks: A vertically integrated HPC framework for elastic FWI

Abstract Elastic full-waveform inversion (FWI) provides high-resolution subsurface models by accounting for pressure- and shear-wave propagation. The multiparameter nature of elastic physics can impose extreme computational demands on the FWI. Modern multicomponent seismic surveys produce terabytes of data and require simulations over billions of grid points, making large-scale elastic FWI challenging even on advanced high-performance computing systems. Although graphics processing units (GPUs) have improved bandwidth and parallelism, suboptimal utilization wastes costly compute cycles and increases project budgets. Focusing solely on a fast computation kernel while neglecting factors such as initialization, data movement, and inter-GPU communication can result in expensive hardware running far below its full potential. A vertically integrated GPU-based elastic FWI framework that addresses the full computational pipeline was developed. By treating performance as a hardware-software codesign problem, key bottlenecks were removed and substantial speedups were achieved without loss of accuracy. This enables faster deterministic inversions and makes advanced uncertainty quantification workflows feasible at realistic survey scales and timelines. Extending these design principles to upcoming specialized processors will enable faster, larger, and more affordable elastic FWI, accelerating its adoption in exploration.