Mapping natural fracture networks using stochastic and machine learning approaches

Traditional stochastic approaches for mapping natural fractures in a reservoir have issues representing the complex connectivity of these networks. A new improved Multiple Point Statistics (MPS) algorithm was tested on the field data obtained from Teapot Dome outcrop fracture image. The algorithm was able to successfully capture the orientations of observed different fracture sets. Our ultimate objective is combination of MPS statistical information with prescribed physical geomechanical criteria for fracture propagation. For this we envision an Artificial Neural Networks (ANN) approach for calibrating the probability of fracture propagation from geomechanics. To that end, we implemented an ANN approach for fracture modeling that mimics the MPS approach. Preliminary results indicate that an ANN by itself may have trouble predicting fracture connectivity accurately due to the limited size of the calibration training set. However, the combination of MPS statistics based fracture modeling and ANN for injecting physics into fracture simulation maybe feasible and that is the future direction of our research.