Conditioning well data to rule-based lobe model by machine learning with a generative adversarial network

• Main Authors: Honggeun Jo, Javier E Santos, Michael J Pyrcz
• Format: Article
• Language: English
• Published: SAGE Publishing 2020-11-01
• Series: Energy Exploration & Exploitation
• Online Access: https://doi.org/10.1177/0144598720937524

Rule-based reservoir modeling methods integrate geological depositional process concepts to generate reservoir models that capture realistic geologic features for improved subsurface predictions and uncertainty models to support development decision making. However, the robust and direct conditioning of these models to subsurface data, such as well logs, core descriptions, and seismic inversions and interpretations, remains as an obstacle for the broad application as a standard subsurface modeling technology. We implement a machine learning-based method for fast and flexible data conditioning of rule-based models. This study builds on a rule-based modeling method for deep-water lobe reservoirs. The model has three geological inputs: (1) the depositional element geometry, (2) the compositional exponent for element stacking pattern, and (3) the distribution of petrophysical properties with hierarchical trends conformable to the surfaces. A deep learning-based workflow is proposed for robust and non-iterative data conditioning. First, a generative adversarial network learns salient geometric features from the ensemble of the training rule-based models. Then, a new rule-based model is generated and a mask is applied to remove the model near local data along the well trajectories. Last, semantic image inpainting restores the mask with the optimum generative adversarial network realization that is consistent with both local data and the surrounding model. For the deep-water lobe example, the generative adversarial network learns the primary geological spatial features to generate reservoir realizations that reproduce hierarchical trend as well as the surface geometries and stacking pattern. Moreover, the trained generative adversarial network explores the latent reservoir manifold and identifies the ensemble of models to represent an uncertainty model. Semantic image inpainting determines the optimum replacement for the near-data mask that is consistent with the local data and the rest of the model. This work results in subsurface models that accurately reproduce reservoir heterogeneity, continuity, and spatial distribution of petrophysical parameters while honoring the local well data constraints.