Uncertainty Quantification in Classifying Complex Geological Facies Using Bayesian Deep Learning

Abstract

Understanding geological differences in a proved reservoir requires precise facies classification. Predicting facies from seismic data is frequently seen as an inverse uncertainty quantification problem in seismic reservoir characterization. Typically, the uncertainty in the model parameters that regulate the geographic distributions is being ignored. The target facies and its uncertainty can be determined by calculating the posterior distribution of a Bayesian inverse problem conditioned by the seismic data, in which the model parameters are inferred from the observed seismic data using a Bayesian inference framework. It is believed that such facies classification model has a unique set of model parameters that best fits it. The proposed work is unique in that it quantifies the epistemic uncertainty of the predicted facies in blind well conditioned by Seismic Amplitude Versus Offset (AVO-Seismic) attributes in the Bayesian inference framework. Under this framework, parameter uncertainties of the neural net. weights and biases are calculated using their posterior distributions from the ensamble models generated by Marcov-Chains Monte-Carlo (MCMC) by assuming that the prior values of the weights and biases are uninformative. The proposed approach is also demonstrated on Synthetic Amplitude Versus Offset (AVO-Synthetic) dataset (derived from the well log information) and we have found high relevance in the predicted results. For comparision, a plain Deep Learning and Deep learning with Monte Carlo Dropout are employed and the results indicate that our model performs more efficiently comparing to the others indicating the possibility of the model to be used in real world solution to adequate facies classication.