Abstract
Ensemble data assimilation methods have been applied with remarkable success in several real-life history-matching problems. However, because these methods rely on Gaussian assumptions, their performance is severely degraded when the prior geology is described in terms of complex facies distributions. This fact motivated an intense investigation reported in the literature to develop efficient and robust parameterizations. Despite the large number of publications, preserving plausible geological features when updating facies models is still one of the main challenges with ensemble-based history matching.
This work reports our initial results towards the development of a robust parameterization based on Deep Learning (DL) for proper history matching of facies models with ensemble methods. The process begins with a set of prior facies realizations, which are used for training a DL network. DL identifies the main features of the facies images, allowing the construction of a reduced parameterization of the models. This parameterization is transformed to follow a Gaussian distribution, which is updated to account for the dynamic observed data using the method known as ensemble smoother with multiple data assimilation (ES-MDA). After each data assimilation, DL is used to reconstruct the facies models based on the initial learning. The proposed method is tested in a synthetic history-matching problem based on the well-known PUNQ-S3 case. We compare the results of the proposed method against the standard ES-MDA (with no parameterization) and another parameterization based on principal component analysis.