Deep Learning Tomography by Mapping Full Seismic Waveforms to Vertical Velocity Profiles

Building realistic and reliable models of the subsurface is the primary goal of seismic imaging. By employing an ensemble of convolutional neural networks (CNNs), we build velocity models directly from seismic pre-stack data and quantify model uncertainties by analyzing all ensemble results. Most attempts are made to infer models as a whole. Here, CNNs are trained to map subsets of seismic data directly into 1D vertical velocity logs. This allows us to integrate well data into the inversion and to simplify the mapping by using the regularity of active seismic acquisition geometries. The presented approach uses neighboring common midpoint gathers (CMPs) for the estimation of individual vertical velocity logs. Trained on augmentations of the Marmousi model, our CNNs allow for the inversion of sections of the Marmousi II and the Overthrust models. Once the ensemble is trained on a particular dataset, similar datasets can be inverted much faster than with conventional full-waveform inversion.