High Resolution Seismic Inversion with Gpu Accelerated TV Regularized Stochastic Lift and Relax Waveform Inversion

Abstract

In seismic exploration, Full Waveform Inversion (FWI) is a crucial tool for imaging subsurface structures. However, it often contends with the persistent challenge of "cycle-skipping," leading to frequent convergence to local minima. This study borrows the idea of model expansion and introduces a novel approach called Lift and Relax Waveform Inversion (LRWI). By lifting unknown variables (wavefields and model parameters) and relaxing the wave-equation constraint, LRWI expands the search space to mitigate the local minima issue. To further accelerate this inversion method, we introduce stochastic optimization with total variation regularization, resulting in TV-Stochastic LRWI (TV-SLRWI). By randomly sampling the data and wavefields and applying TV regularization to the model parameters, we achieve over a ten-fold acceleration in the inversion process while maintaining inversion accuracy. Additionally, we have implemented a parallelization scheme with Graphics Processing Unit (GPU) acceleration. Numerical results demonstrate that the proposed method effectively overcomes the local minima problem of conventional FWI. Furthermore, the GPU-accelerated parallelization scheme provides an impressive 210-fold speedup compared to Central Processing Unit (CPU)-based implementations.

Keywords: full waveform inversion, seismic imaging, GPU, Stochastic Optimization, LRWI