Massive Geomodel Compression and Rapid Geomodel Generation Using Advanced Autoencoders and Autoregressive Neural Networks

Abstract The reduction of computational cost when using large geomodels requires low-dimensional representations (transformation or reparameterization) of large geomodels, which need to be computed using fast and robust dimensionality reduction methods. Additionally, to reduce the uncertainty associated with geomodel-based predictions, the probability distribution/density of the subsurface reservoir needs to be accurately estimated as an explicit, intractable quantity for purposes of rapidly generating all possible variability and heterogeneity of the subsurface reservoir. In this paper, we developed and deployed advanced autoencoder-based deep-neural-network architectures for extracting the extremely low-dimensional representations of field geomodels. To that end, the compression and reconstruction efficiencies of vector-quantized variational autoencoders (VQ-VAE) were tested, compared and benchmarked on the task of multi-attribute geomodel compression. Following that, a deep-learning generative model inspired by pixel recurrent network, referred as PixelSNAIL Autoregression, learns not only to estimate the probability density distribution of the low-dimensional representations of large geomodels, but also to make up new latent space samples from the learned prior distributions. To better preserve and reproduce fluvial channels of geomodels, perceptual loss is introduced into the VQ-VAE model as the loss function. The best performing VQ-VAE achieved an excellent reconstruction from the low-dimensional representations, which exhibited structural similarity index measure (SSIM) of 0.87 at a compression ratio of 155. A hierarchical VQ-VAE model achieved extremely high compression ratio of 667 with SSIM of 0.92, which was further extended to a compression ratio of 1250 with SSIM of 0.85. Finally, using the PixelSNAIL based autoregressive recurrent neural network, we were able to rapidly generate thousands of large-scale geomodel realizations to quantify geological uncertainties to help further decision making. Meanwhile, unconditional generation demonstrated very high data augmentation capability to produce new coherent and realistic geomodels with given training dataset.