Abstract: |
Multiple-point statistics are widely used to simulate complex heterogeneous fields. The technique consists in inferring multiple-point statistics of a categorical variable from a training image. For 3D problems with numerous facies, large templates should be used for reproducing complex structures properly. Tree structures used in classical implementations for storing the multiple-point statistics of the training image are very RAM demanding and then imply limitations on the size of the template. On the contrary, using a list structure allows to obtain a straightforwardly parallelizable algorithm requiring a small amount of RAM. Nevertheless, retrieving statistics from a search tree is more efficient in terms of CPU time, because the shortcuts given by the branches of the tree are not present in the list. In this paper, we propose a new technique mixing list and tree structures for storing multiple-point statistics inferred from the training image. The idea is to build a tree of reduced size whose leaves are sub-lists that constitute the entire list when gathering them. This approach benefits from the advantages of both storage techniques: low RAM requirements are guaranteed by the list structure, while improved efficiency in terms of CPU time is provided by the tree structure and the parallelization. Numerical tests are performed for comparing the different methods and presented in this paper. |
