Simplified Dynamic Modeling of Faulted Turbidite Reservoirs: A Deep-Learning Approach to Recovery-Factor Forecasting for Exploration

Abstract

Summary Faults are geological features that are essential in considering the development of hydrocarbon reservoirs. Significant resources (appraisal costs and manpower) are often deployed to locate them and assess their connectivity more accurately so that adequate investment decisions can be made. Early in a project's life, data might be sparse or the time that can be dedicated to data processing might be limited. An estimate of the impact of faults can be derived by considering the impact of analogous fault patterns in similar reservoir environments. A significant number of discoveries consist of channelized turbidite reservoirs draped over deepwater toe–thrust anticlines. To understand the effects of various fault patterns on the recovery factors of structurally complex turbidite reservoirs, we first perform extensive flow–simulation–based multidimensional sensitivity studies using realistic channelized stratigraphic architectures. Many of these reservoirs contain high–porosity, high–permeability channels, light oil, limited aquifer drive, and a limited number of fault populations. To constrain the potential impacts of the faults, a generic reservoir model was constructed and simulated with varying reservoir and fault properties. Because the input parameters are known (e.g., fault–zone thickness and permeability, shale–drape coverage, and oil viscosity), the relative importance of each variable can be quantified. The simulation results show that the impacts of the faults on reservoir performance vary systematically with the fault length and orientation, the undeformed–reservoir permeability, and the fault permeability. The recovery factors are relatively insensitive to fault–zone thickness, net/gross ratio, and porosity. The time–delay effect of the faults is significant in reservoirs with a permeability of approximately 100 md but not in darcy–quality reservoirs. In most fault scenarios, the recovery factors from simulations with high–permeability faults range from approximately 30% in 100–md reservoirs to approximately 45% in 3,000–md reservoirs, equivalent to 75 to 96% of the recovery factors of the unfaulted reservoirs. In comparison, assuming the most likely fault permeabilities, the recovery factors decrease to circa 20% in 100–md reservoirs and circa 41% in 3,000–md reservoirs, equivalent to 40 to 85% of the values in unfaulted reservoirs. To learn and predict the effects of faults on the recovery factor in a very fast and reasonably generic fashion in deepwater toe–thrust anticlines, we have developed an accurate deep–learning (DL) based surrogate model that predicts the structure connectivity factors as a function of recovery time. These factors are then used for discounting the recovery factors computed by simpler and easier–to–construct unfaulted models. The surrogate model is a very fast and advanced proxy for the multidimensional structure connectivity factor function and operates without resorting to any new flow simulation with a faulted model. We describe the novel DL architecture used for constructing the surrogate model and quantitatively prove its effectiveness. We finally demonstrate an example real–life application of the surrogate model on an exploration–stage prospect–ranking exercise.