Unconfined deep-water lobe deposits are among the most important targets in deep-water oil field exploration and production. Accurate stochastic simulations of the sedimentary architectures and petrophysical properties of deep-water lobe deposits require robust seismic and well data integration. The reservoir heterogeneity at scales below seismic resolution generally exhibits important and predictable hierarchical architectures that control the vertical and horizontal connectivity of the reservoir, affecting hydrocarbon recovery rate during development. Current geostatistical simulation algorithms such as variogram– and multiple-point–based methods readily perform property modeling conditioned to well data and trends informed from seismic data at and above seismic resolution. Yet, these common geostatistical models are limited in their ability to reproduce essential, multiscale heterogeneities below seismic resolution between wells, including nested, multiscale architectures and trends. Rule-based methods are commonly used for modeling the hierarchical architectures, but conditioning the models to well logs and seismic data is still limited, difficult, and time-consuming. Either a prohibitive degree of expert intervention is employed resulting in overly deterministic models or the common “cookie-cutter” approach is applied, resulting in inconsistency over the multiple scales.

To address the well and seismic data conditioning and multiscale modeling limitations of current geostatistical modeling methods, we propose a new workflow based on the hierarchical application of the newly developed stochastic pix2pix machine learning algorithm. The hierarchical stochastic pix2pix workflow allows us to extract stacking and geometrical patterns of architectural elements in deep-water lobe deposits by learning from rule-based training models, and stochastically calculate a diverse ensemble of conditional, multiscale, architectural and petrophysical, property model realizations reproducing the patterns. We demonstrate the hierarchical stochastic pix2pix workflow to calculate an ensemble of multiscale, deep-water lobe depositional systems conditioned by (1) geostatistical rule-based training and testing models, and (2) bounding surfaces and property estimates from seismic and well log interpretations.

With the hierarchical stochastic pix2pix workflow, we can efficiently and accurately calculate diverse three-dimensional reservoir models, which are conditional to the well-log and seismic interpretations and reproduce multiscale heterogeneity. We also use quantitative measures, such as a raster-based compensational index and the dynamic Lorentz coefficient to validate the model accuracy. In addition, the reduced model parametric representation and efficient model calculation available with the hierarchical stochastic pix2pix workflow is critical to the practical solutions of reservoir inverse problems.