AUTOMATIC UNFLOODING FOR SALT BASE USING U-NET IN FULL-WAVEFORM INVERSION FRAMEWORK

Velocity model building in salt-affected regions is a major challenge. The long-standing industry practice consists of picking the top/base of the salt from seismic images for flooding/unflooding the salt velocity. The bottom of the salt is often unclear and difficult to pick, even by experts. Machine learning can overcome human limitations in pattern recognition, and thus, to recognize the base of the salt. In a supervised learning framework, we generate many 1D models containing flooded salt bodies and invert for their velocity using FWI. Then, we use the inversion results as input and the true model as labels to train the network to unflood the velocity to the correct depth. After training, the neural network takes the vertical profile from 2D models by FWI and outputs a model automatically unflooded. We show the potential of the trained network on the west part of BP 2004 salt model. We will show real data applications in the presentation.