A graphics processing unit implementation of time-domain full-waveform inversion |
We have implemented GPU-based FWI using the wavefield reconstruction strategy, which averts the issue of CPU-GPU data transfer. The Clayton-Enquist absorbing boundary was utilized to maintain the efficiency of GPU computation. A hybrid nonlinear conjugate gradient method combined with parallel reduction technique was adopted in the FWI optimization. The validity of our implementation for GPU-based FWI was demonstrated using a numerical test.