Machine learning assisted history matching for a deepwater lobe system

Abstract

High exploration costs resulting in sparse datasets and complicated geological structures in deepwater depositional systems make the reservoir characterization extremely difficult. To meet this challenge, rule-based geostatistical subsurface modeling has been developed to fill this data gap with enhanced integration of geological concepts for deepwater reservoirs based on geological rules constrained by this sparse well data. The rule-based model simulates sediment dynamics through depositional rules to calculate reservoir architecture and the associated rock properties distributions. As a result, rule-based models incorporate conceptual and qualitative information, such as temporal deposition sequence and consequent compensational stacking patterns. Integrating quantitative data, such as fluid production history, is a remaining obstacle to the broad application of rule-based subsurface models. We propose a new machine learning assisted history matching workflow for rule-based models. First, multiple rule-based models are calculated as training data for a Generative Adversarial Network (GAN). The successfully trained GAN enables exploration of the latent reservoir manifold with only a small number of numeric values (i.e., latent random vector); this is a massive reduction in the dimensionality of potential subsurface models. The initial ensemble models, generated by the trained GAN with a range of latent random vectors, are coupled with physics-based reservoir flow simulation to obtain production forecasts. Ensemble Kalman filter (EnKF) updates the latent random vectors of the ensemble by minimizing the misfit (i.e., error norm) between the production forecasts and the production observation over time. Our proposed workflow finds an ensemble of optimized reservoir models that honor realistic geological heterogeneity and production history. Besides, as the GAN's latent random vectors are Gaussian distributed, one of the fundamental mathematical assumptions of EnKF, this workflow effectively alleviates possible artifacts (e.g., filter divergence or overshooting) in EnKF. Moreover, this proposed workflow is more computationally efficient than updating the entire high-dimensional reservoir properties and removes the limitation to only Gaussian simulation models with their limited geological realism. The proposed workflow may be expanded to various reservoir depositional settings. • We propose a machine learning assisted history matching workflow for complex deepwater subsurface models. • We apply generative adversarial networks (GAN) to represent the complex subsurface models by latent random vectors. • We update latent random vectors of the subsurface model to match the production history by the ensemble Kalman filter (EnKF). • By combining GAN and EnKF, we can 1) enhance computational efficiency, 2) remove artifacts, and 3) preserve geologic realism.