Distance Based Simulation for Generating Additional Realizations

Running dynamic data integration algorithms such as sequential self-calibration requires a large amount of CPU time. It has been observed that not all the resulting realizations support the desired level of measurement error; and therefore a screening is required as a post-processing step. The required screening adds to the time needed for generating the desired number of realizations. An alternative way of expanding the set of Gaussian realizations supporting a particular level of measurement error has been examined in this work. This alternative approach, termed distance-based simulation (Scheidt et al. 2008 and Caers 2008), builds on the concepts of multi-dimensional scaling, KL-expansion, kernel principal component analysis and modeling and simulation in metric and feature spaces. It is observed that this technique needs a very careful post-processing of the realizations to ensure the reproduction of input histogram (especially standard deviation) by the expanded set of realizations. It is also observed that many of the generated realizations may closely resemble the input realizations.