Bayesian variational auto-encoder for seismic wavelet extraction

Seismic wavelet extraction is a crucial component of processes relating seismic data to reflectivity series, such as seismic-well-tie or seismic inversion. We propose a parameter-free Bayesian Variational Auto-Encoder (BVAE) wavelet extractor trained in a supervised manner on synthetically generated training data. To achieve that, we feed the network with pairs of synthetic reflectivity series and seismic, labelled by a corresponding wavelet and utilize a Bayesian network architecture to capture the uncertainty of the extraction process. The goal of this deep-learning based approach is to minimize the amount of required parameters, stabilize the extraction process in the limited well-log data scenario and provide uncertainty ranges for both, the frequency and phase of the estimated wavelet. Moreover, we demonstrate the capability of the model to capture uncertainty in a synthetic case, where ground-truth is present and demonstrate its usability in a real post-stack well-tie example.