Mapping full seismic waveforms to vertical velocity profiles by deep learning

Abstract

Building realistic and reliable models of the subsurface is the primary goal of seismic imaging. Here we construct an ensemble of convolutional neural networks (CNNs) to build velocity models directly from the data. Most other approaches attempt to map full data into 2D labels. We exploit the regularity of seismic acquisition and train CNNs to map gathers of neighboring common midpoints (CMPs) to vertical 1D velocity logs. This allows us to integrate well-log data into the inversion, simplify the mapping by using the 1D labels, and accommodate larger dips relative to using single CMP inputs. We dynamically generate the training data in parallel with training the CNNs, which reduces overfitting. Data generation and training of the CNNs is more computationally expensive than conventional full-waveform inversion (FWI). However, once the network is trained, data sets with similar acquisition parameters can be inverted much faster than with FWI. The multiCMP CNN ensemble is tested on multiple realistic synthetic models, performs well, and was combined with FWI for even better performance.

Quantification of DAS VSP Quality: SNR vs. Log-Based Metrics

“…Next, the log-based kinematic and dynamics metrics are calculated. They include the mean absolute percentage error (MAPE), the normalized root-mean-square deviation (NRMSD), and the R2 score [20], which we define in the next section (Equations ( 3)-( 5)).…”

Adaptive Feedback Convolutional‐Neural‐Network‐Based High‐Resolution Reflection‐Waveform Inversion

“…Kazei et al. (2021) propose to generate random velocity models from a guiding model, by applying complicated image processing procedures (flipping, cropping, distortion, etc.). Then, the CNN is trained to approximate the mapping from full seismic waveforms to 1D vertical velocity profiles, which, compared to a 2D velocity model, are more randomly sampled and selected (with less selection‐bias).…”

Coupling Deep Learning with Full Waveform Inversion

“…In recent years, there are great interest in the FWI community to use deep learning techniques, based on neural networks, to replace the classical least-squares based inversion methods [1,4,16,22,23,28,31,32,36,39,49,54,55,56,57,63,65,66,67,68]. Assume that we are given a set of sampled data…”

“…Numerical experiments, such as those documented in [4,32,36,54,63,65,66,67,68], showed that, with sufficiently large training datasets, it is possible to train highly accurate inverse operators that can be used to directly map measured wave field data into the velocity field. This, together with the recent success in learning inverse operators for other inverse problems (see for instance [2,7,11,24,47,53] for some examples) has led many to believe, probably overly optimistically, that one can completely replace classical computational inversion with offline deep learning.…”