Enhancing Forecasting Accuracy Through Reservoir Modeling with Computational Stratigraphy (CompStrat) and Hierarchical Generative Adversarial Networks (GAN)

Abstract Accurate subsurface modeling significantly enhances forecasting performance, providing more precise predictions of reservoir behavior and improved decision making. Traditional geostatistical methods, such as variogram models and multiple point statistics, honor hard data but struggle with complex heterogeneity and non-stationary distributions, leading to unrealistic facies distribution. Conversely, realistic physics-based models including digital analogs, like high-resolution shallow seismic and lidar-scanned outcrops, face challenges when integrating hard data. To address these challenges, we propose a machine learning-based approach called Machine Learned Stratigraphy (MLS) for facies reservoir modeling. MLS leverages hierarchical generative adversarial networks (GAN) trained on digital analogs from state-of-the-art Computational Stratigraphy technology to generate geologically realistic, data-conditioned facies models across multiple scales. It employs a hierarchy of GANs, each conditioned in the latent space to hard and soft data with varying support sizes (seismic, well logs, and petrophysical), facilitating easier data conditioning. The trained model generates multiple equiprobable realizations with realistic heterogeneities and varying levels of connectivity. This method significantly improves the reliability of the reservoir model by accurately representing subsurface complexities. It also improves forecasting accuracy through numerical flow simulation, leading to better predictions of reservoir behavior and performance. Compared to traditional geostatistical and machine learning methods, this technique is faster, more resource-efficient and avoids common modeling artifacts.