Semi-Supervised Seismic Facies Prediction via Gaussian Mixture and Variational Models

The absence of labeled data complicates the application of supervised machine learning algorithms. So, curating labeled training data has become the primary bottleneck in machine learning. Deep generative models have been proposed to synthesize labels at a proper scale in areas with weak supervision sources to accomplish the labeled data scarcity problem. Recently, a joint labeled seismic data expansion generative method has been proposed based on Variational Autoencoders and Gaussian Mixture Models. In this paper, we extend this strategy by using as input multi-channel data (pre-stack full-azimuth seismic data) as considering independent channels by azimuths. Moreover, a Bayesian Gaussian Mixture Model prior conditioned by the variational inference is proposed to fit the deep feature distribution of each class. The probabilistic Gaussian mixture model is resampled for each class to provide depth features expansion into the decoder and generate expansion-labeled seismic data. This strategy is applied to a Santos basin Pre-salt reservoir to expand labeled mounds facies identified by wells. The approach quickly overcame an important labeled data issue to support seismic characterization, minimizing overfit problems and improving the recognition of mounds’ architectural elements in the field.