Summary
Efficient identification and characterization of fracture networks are crucial for the exploitation of fractured media such as naturally fractured reservoirs. Using the information obtained from borehole logs, core images, and outcrops, fracture geometries can be roughly estimated. However, this estimation always has uncertainty, which can be decreased using inverse modeling. Following the Bayes framework, a common practice for inverse modeling is to sample from the posterior distribution of uncertain parameters, given the observational data. However, a challenge for fractured reservoirs is that the fractures often occur on different scales, and these fractures form an irregular network structure that is difficult to model and predict. In this work, a multiscale-parameterization method is developed to model the fracture network. Based on this parameterization method, we present a novel history-matching approach using a data-driven evolutionary algorithm to explore the Bayesian posterior space and decrease the uncertainties of the model parameters. Empirical studies on hypothetical and outcrop-based cases demonstrate that the proposed method can model and estimate the complex multiscale-fracture network on a limited computational budget.
Physics constrained learning for data-driven inverse modeling from sparse observations
