Deep-learning seismic facies on state-of-the-art CNN architectures

In the 1950s neural networks started as a simple direct connection of several nodes in an input layer to several nodes in an output layer (Widrow and Lehr, 1990). In geophysics this puts us to the introduction of seismic trace stacking (Yilmaz, 2001). In 1989 the first idea of a convolutional neural network was born (Lecun, 1989) and back-propagation was formalized as an error-propagation mechanism (Rumelhart et al., 1988). In 2012 the paper (Krizhevsky et al., 2012) propelled the field of deep learning forward implementing essential components, namely GPU training, ReLu activation functions (Dahl et al., 2013) and dropout (Srivastava et al., 2014). They outperformed previous models in the ImageNet challenge (Deng et al., 2009) by almost halving the prediction error. Waldeland and Solberg (2016) showed that neural networks can be used to classify salt diapirs in 3D seismic data. Charles Rutherford Ildstad (2017) generalized this work to nD and beyond two classes of salt and ”else”.