Stochastic inversion of discrete fracture networks using genetic algorithms

Abstract

Reservoir management and contaminant transport simulations rely on accurate modeling of the subsurface. This task becomes more challenging when the reservoir of interest also has a large amount of fractures. Inverse modeling of these fractured media involves first performing multiple field tests such as pumping tests, tracer tests, or seismic surveys. Inverse modeling methods then are used to find potential geologic models that produce simulated results that match field observations. Popular methods include the Ensemble Kalman Filter, Markov Chain Monte Carlo, and simulated annealing. A common challenge these methods face is that inverse modeling of fractured media is inherently a multiobjective optimization task. The inverse modeling method must find a discrete fracture model that produces the same flow characteristics as observed in field pumping or tracer tests. But it also must find a discrete fracture model with a fracture network that matches the fracture parameter distribution observed by the field measurements such as seismic surveys. This challenge can be approached in two steps. The first step is producing discrete fracture network models with parameter distributions that match field observations from seismic surveys. This study focused on the second step, involving the development of a method that can take a population of discrete fracture networks and generate new fracture networks in a way that preserves the fracture parameter distribution. This was done using a genetic algorithm modified to apply to the domain of discrete fracture networks. During genetic mixing, the fractures of the child model are generated by randomly copying over fractures from the parent models. This process ensures the child model adopts the fracture parameter distribution of the parent models. This study tests the effectiveness of this genetic algorithm on a synthetic example. The results of the experiment show the genetic algorithm is able to effectively produce a population of discrete fracture models with breakthrough curves that match the curves of the reference model.