next up previous [pdf]

Next: FWI and its GPU Up: Yang et al.: GPU Previous: Yang et al.: GPU

Introduction

The classical time-domain full waveform inversion (FWI) was originally proposed by Tarantola (1984) to refine the velocity model by minimizing the energy in the difference between predicted and observed data in the least-squares sense (Symes, 2008). It was further developed by Tarantola (1986) with applications to elastic cases (Pica et al., 1990). After Pratt et al. (1998) proposed frequency domain FWI, the multiscale inversion became an area of active research, and provided a hierarchical framework for robust inversion. The Laplace-domain FWI and the Laplace-Fourier domain variant have also been developed by Shin and Cha (2009,2008). Until now, building a good velocity model is still a challenging problem and attracts increasing effort of geophysicists (Virieux and Operto, 2009).

There are many drawbacks in FWI, such as the non-linearity, the non-uniqueness of the solution, as well as the expensive computational cost. The goal of FWI is to match the synthetic and the observed data. The minimization of the misfit function is essentially an iterative, computationally intensive procedure: at each iteration one has to calculate the gradient of the objective function with respect to the model parameters by cross correlating the back propagated residual wavefield with the corresponding forward propagated source wavefield. The forward modeling itself demands large computational efforts, while back propagation of the residual wavefield has large memory requirements to access the source wavefield.

Recent advances in computing capability and hardware makes FWI a popular research subject to improve velocity models. As a booming technology, graphics processing unit (GPU) has been widely used to mitigate the computational drawbacks in seismic imaging (Yang et al., 2014; Micikevicius, 2009) and inversion (Shin et al., 2014; Boonyasiriwat et al., 2010), due to its potential gain in performance. One key problem for GPU implementation is that the parallel computation is much faster while the data communication between host and device always takes longer time. In this paper we report a 2D implementation of GPU-based FWI using a wavefield reconstruction strategy. The boundaries of the forward modeling are saved on the device to avert the issue of CPU-GPU data transfer. Shared memory on the GPU is used to speedup the modeling computation. A hybrid nonlinear conjugate gradient method is adopted in the FWI optimization. In each iteration, a Gaussian shaping step is employed to remove noise in the computed gradient. We demonstrate the validity and the relatively superior speedup of our GPU implementation of FWI using the Marmousi model.


next up previous [pdf]

Next: FWI and its GPU Up: Yang et al.: GPU Previous: Yang et al.: GPU

2021-08-31