VelocityGAN: Subsurface Velocity Image Estimation Using Conditional Adversarial Networks
Highlight
The usual approach to alleviate ill-posedness is to incorporate prior knowledge with a regularization term that penalizes solutions that are inconsistent with this prior knowledge. Most existing regularization techniques employ generic functions (e.g., L1 or L2 penalties on coefficients) that are loosely (if at all) related to the physical problem at hand. We implement a data-driven inversion method as an image-to-image-translation problem using generative adversarial networks (GANs) structure, which I call “VelocityGAN”. VelocityGAN learns an effective regularization that is customized to the inversion problem. In particular, we use GANs to learn a classifier to discriminate between the true and the generated velocity maps. The discriminator penalizes velocity maps that do not “look like” the maps that are used for training. We compare the performance of VelocityGAN with InversionNet and notice some improvement. Most importantly, we notice that VelocityGAN yields generalization ability to some extent,
The full paper can be found on arxiv:
Data-driven Seismic Waveform Inversion: A Study on the Robustness and Generalization
Results
Illustration of some examples from SEG/EAEG salt data set. We show six inversion results from testing set to demonstrate that our model can predict velocity maps with varying salt dome shapes. For each example, the ground truth image is shown on the left column and the inverted image is shown on the right column. The colormap used ranges from the minimum value to maximum value of each pair of velocity maps.
