Autoencoders for seismic model upscaling and facies identification

Access full-text files




DeFabry, Cameron Mark

Journal Title

Journal ISSN

Volume Title



The research presented here focuses on the resolution enhancement of inverted seismic volumes and geological facies identification. In the first section I utilize a regularized neural network in the form of an autoencoder to improve the resolution of seismic models which were inverted for compressional wave impedance. In the second section, I focus on the utilization of autoencoders to classify spatially small geologic facies with inverted seismic models and a computed facies map. In section one I focus on the processing and inversion of seismic data from the 3D Penobscot field on the Scotian shelf to produce models for compressional wave impedance. The dataset is inverted using two different approaches, first using a deterministic method, and secondly using a stochastic method. The stochastically derived models have higher frequency content than the deterministic models and are used as the target for the task of resolution enhancement. An autoencoder is trained to recreate the stochastic models with a set of randomly chosen starting weights. During training the network attempts to create a feature map that correlates the low-resolution deterministic model to the high-resolution stochastic model by using the deterministic models as input data and the stochastic models as target data. Once training is complete the network is given the deterministic model as an input and asked to predict an output with the convolution filters learned by recreating the stochastic models. The result is a model of higher resolution than the original deterministic model, but lower resolution than the original stochastic model. In section two I characterize a seismic volume from the Marco Polo field collected in the Gulf of Mexico and then classify five distinct facies, a shale member, an oil sand member, a gas sand member, and discrete brine sand members corresponding to the sand units. The brine sand members were simulated through fluid substitution and then have their probabilistic properties derived through a rock physics template. Bayesian classification is used to create an initial facies map with the brine sands predicted with rock physics templates, and the remaining units predicted directly from the distributions inherent to the well log. Geologic units less than 600 meters in any direction were specifically targeted by converting the standard facies map into a binary facies map. This binary facies map was used as an input in an autoencoder along with two seismic volumes inverted for compressional wave impedance and Vp/Vs. The result is a trained network that can take inverted model inputs and produce a probabilistic output predicting the location of a given facies. Additionally, when provided with smoothed inputs, the autoencoder can produce outputs of a similar resolution to that of the original data, with a loss of performance noted in the probabilities displayed. This result, along with the result from section one are used to justify the claim that autoencoders can be effectively used for the tasks of seismic model upscaling and facies identification without the direct use of well log data as a network input. Convolutional layers provide a way of processing these data in a manner often seen in image recognition and enhancement problems. These networks are limited to the data they were trained on, so additional training would be required for use on separate datasets. The utilized methods though, should maintain their efficacy provided the appropriate training has taken place.


LCSH Subject Headings