CONDITIONING RULE-BASED MODEL TO STRATIGRAPHY WITH MACHINE LEARNING: DEMONSTRATION IN DEEPWATER LOBE SYSTEM

The stratigraphic rule-based models approximate sedimentary dynamics to generate realistic subsurface architectures. A few intuitive rules for the sequential depositional units render realistic reservoir heterogeneity to petrophysical property distributions. However, due to a significant degree of uncertainty in geological history (e.g., sea-level change, sediment supply, and preservation), rule-based models often rely on the stochastic application of rules over the sequence of depositional units. This sequential, stochastic approach for rule-based models reduces the conditioning of these rule-based models to the stratigraphy over the observed scale. Stratigraphy is the container of the geological event, and heterogeneity indicates the inner structure in the container.

This study proposes a machine learning assisted workflow for conditioning the rule-based models to stratigraphy. First, we generate multiple rule-based models as a training dataset. Then, we train a Generative Adversarial Network (GAN) to learn the intrinsic geological features, such as hierarchical structure, element geometry, and stacking rule. On top of latent variables (i.e., standard inputs for GANs), we add the profile of the compositional surfaces of the rule-based models to the input of the GAN. As such, with the trained GAN, we can generate a new subsurface model, which honors the given stratigraphy with the primary realistic geological feature. Finally, the trained GAN takes the stratigraphy profile from observation (e.g., outcrop analogy or seismic data) and generates the subsurface models that are conditioned to the observation.

We demonstrate our workflow in the deepwater lobe system. This workflow results in a suite of subsurface models that honor both the realistic heterogeneity and observed stratigraphy. Moreover, albeit we only use simple lobe geometry in generating the training dataset, the GAN successfully adjusts the shape and size of the lobe to be consistent with the local stratigraphy. As such, this workflow enables the reproduction of more complicated geological features, such as variation in the sediment flow direction and sediment erosion, consistent over multiple scales for improved subsurface modeling accuracy and uncertainty assessments to support optimum development decision making.