Author(s): |
O. Peredo, Barcelona Supercomputing Center (ES)
J. M. Ortiz, ALGES Lab, Advanced Mining Technology Center, University of Chile (CL)
|
Abstract: |
Multiple-point geostatistical simulation methods allow building numerical models that reproduce pattern statistics conditioned to hard data. These models are used for uncertainty quantification purposes in diverse applications of geological characterization, in areas such as geological resource estimation, oil reservoir modeling, and hydrogeological forecasting. Genetic algorithms were developed to solve combinatorial optimization problems, and as such, are suited to be used in the construction of these models, by starting with a model that does not comply with the required statistics and by successively applying the genetic operators, in order to approach the statistics to the target ones, which are inferred from a training image. The procedure consists on defining a fitness function, associated to the mismatch between the current model statistics and the target statistics for a pattern of a given size, and selecting the best individuals (realizations). A crossover and mutation process is then applied, where the selected individuals are combined and then some of their nodes are changed. These new individuals (realizations) replace the old ones and a new state is achieved. The repeated application of this process allows improving (reducing) the average fitness function value, converging towards the target statistics, inferred from a training image. The procedure is demonstrated considering several patterns, and its implementation in a shared-memory supercomputer is discussed. Results show that a good reproduction of the imposed statistics can be achieved and the algorithm has a large potential for improving its performance. |
