Unsupervised Seismic Facies Deep Clustering Via Lognormal Mixture-Based Variational Autoencoder

Seismic facies analysis (SFA) is a crucial step in the interpretation of subsurface structures, with the core challenge being the development of automatic approaches for the analysis of 4D prestack seismic data. The dominant isolated learning-based SFA schemes have gained considerable attention and primarily focus on learning the best representation of prestack data and generating facies maps by clustering the extracted features. However, in isolated learning, the independent nature of feature extraction and clustering leads to the ineffectiveness of clustering loss guidance on feature extraction, thereby resulting in derived features that unnecessarily facilitate the clustering task. As an alternative, we proposed a new unsupervised, end-to-end learning-based SFA method, which is referred to as the lognormal mixture-based variational autoencoder (LMVAE) and enhanced the existing Gaussian mixture variational autoencoder-based deep clustering framework (GMVAE framework). In this approach, both the extraction and clustering of seismic features are simultaneously performed by determining from which mode of the latent mixture distribution the seismic data were generated. Furthermore, the LMVAE extends the Gaussian mixture modeling of seismic features in the GMVAE framework to lognormal mixture modeling, improving the adaptability of SFA to field data. The effective performance of the LMVAE is demonstrated in synthetic and field prestack seismic data.