A hybrid approach for multiresolution modeling of large-scale scientific data

Abstract

Simulations of complex scientific phenomena involve the execution of massively parallel computer programs. These simulation programs generate large-scale multidimensional data sets over the spatio-temporal region. Analyzing such massive data sets is an essential step in helping scientists glean new information. To this end, efficient and effective data models are needed. In this paper, we present a hybrid approach for constructing data models from large-scale multidimensional scientific data sets. Our models not only provide descriptive information about the data but also allow users to subsequently examine the data by querying the data models. Our approach combines a multiresolution-topological model of the data with a multivariate-physical model of the data to generate one hierarchical data model that efficiently captures both the spatio-temporal and the physical aspects of the data. In particular, this hybrid approach consists of three phases. In the first phase, we build a multiresolution model that encapsulates the data set's spatial information (i.e., topology and spatial connectivity). In the second phase, we build a multivariate model from the physical dimensions of the data set. Physical dimensions refer to those dimensions that are neither spatial (x, y, z) nor temporal (time). The exclusion of the spatial-temporal dimensions from the clustering phase is important since similar characteristics could be located (spatially) far from each other. Finally, in the third phase, we connect the multivariate-physical model to the multiresolution-topological model by utilizing ideas from information retrieval. The third phase is essential since the multivariate-physical model does not contain any topological information (without which the model does not have accurate spatial context information). Experimental evaluations on two large-scale multidimensional scientific data sets illustrate the value of our hybrid approach.