Application of Graph Variational Autoencoders for History Matching Problem of Brugge Field

We present an approach to generate reservoir models conditioned on well log and production data along with uncertainty estimation with Variational Autoencoders based on Graph Convolutions. This approach demonstrates the ability to implicitly parameterize geological representations into a latent space of reduced dimensionality and provides ways to uncertainty quantification and production profiling across multiple geological concepts. GVAE links the static model and its dynamic response via latent space, which is a reduced-order representation of geological model complexity. GVAE can handle unstructured data represented as graphs, opposed to a conventional VAE applied in similar works. On the Brugge field dataset, we showed that GVAE reliably reproduces geology by comparison static and dynamic properties of reference (initial) model with generated representations. GVAE utilizes the notion of a latent space depicting the variation of geological concepts and other geological properties. Navigation and optimization through the latent space provide a model update and an ensemble of history-matched models.