Bayesian Convolutional Neural Networks for Seismic Facies Classification

The seismic response of geological reservoirs is a function the elastic properties porous rocks, which depends on rock types, petrophysical features, and environments. Such characteristics are generally classified into facies. We propose to use convolutional neural networks in Bayesian framework predict facies based data quantify uncertainty classification. A variational approach adopted approximate posterior distribution parameters that mathematically intractable. network trained labeled examples. mean standard deviation randomly drawn from predefined Gaussian functions for initialization, updated by minimizing negative evidence lower bound. classification applied sections not included training set. draw multiple random samples simulate an ensemble predictor classify most probable implement proposed open-source library TensorFlow Probability, its convenience flexibility. applications show internal regions with higher confidence than their boundaries, as measured predictive entropy calculated multiclass probability across possible plain also comparison, assigning fixed values using classical backpropagation technique. comparison shows consistent results; however, able assess predictions.