Lab Home | Phone | Search | ||||||||
|
||||||||
Seismic imaging using the full waveform inversion has important applications in reservoir identification and carbon sequestration, however, it requires costly forward modeling. To increase efficiency using neural surrogate models, we identify several techniques motivated by properties of the wave equation for improving solution accuracy in the frequency domain, including an expressive source term encoding and wavefield mixup. We demonstrate improved performance on the OpenFWI dataset. Host: Hanchen Wang (T-5) and Youzuo Lin (EES-17) |