Reservoir Modeling With Deep Learning

Abstract

The parameterization of conventional geologic models for history-matching to reservoir production data is often inconvenient for preserving geologic realism. Each geologic modeling technique requires a unique strategy for calibrating to both geologic well and reservoir production data, and consequently geologic uncertainty is typically ignored for history-matching. A novel reservoir modeling framework utilizing unsupervised deep learning is proposed to learn, reproduce and calibrate any three-dimensional geologic model to geologic well and historical reservoir production data. Firstly, training data representing the geologic model of interest are required. Training data can be created using any geologic modeling technique. Secondly, a Generative Adversarial Network (GAN) is used to stochastically generate realistic geologic models. The GAN is represented by two convolutional neural networks, a Discriminator and a Generator, learning in a game-theoretic approach. After training, the Generator can rapidly map samples from a normally distributed latent input to a three-dimensional geologic model. Thirdly, an additional neural network is trained to condition geologic models to geologic well data by modifying the distribution of the latent input to the Generator. This parameterization of conditioned geologic models is useful for ensemble-based history-matching methods that rely on the Gaussian assumption for the model parameters. In the final step, the conditional geologic models are directly optimized to honor historical reservoir production data using an iterative ensemble smoother. An ensemble of reservoir models honoring geologic well data and historical reservoir production data, inspired by a fluvial reservoir in the North Sea, United Kingdom, is generated. Results are compared to conventional methods of history-matching fluvial reservoir models to demonstrate that the proposed methodology is capable of matching observed reservoir production data and preserving geologic realism. Although a fluvial reservoir example is presented, the framework is generic and facilitates improved forecast reliability by integrating realistic geologic models together with calibration to geologic well and historical reservoir production data.

AAPG Datapages/Search and Discovery Article #90350 © 2019 AAPG Annual Convention and Exhibition, San Antonio, Texas, May 19-22, 2019