Cluster-based Analysis of Multi-Parameter Distributions in Cloud Simulation Ensembles
Department of Informatics, Technische Universität München (TUM), Germany
The proposed approach enables a comparative visual exploration of multi-parameter distributions in time-varying 3D ensemble simulations. To investigate whether dominant trends in such distributions occur, we consider the simulation elements in each dataset—per ensemble member and time step—as elements in the multi-dimensional parameter space, and use t-SNE to project these elements into 2D space. To find groups of elements with similar parameter values in each time step, the resulting projections are clustered via k-Means. Since elements with similar data values can be disconnected in one single projection, we compute an ensemble of projections using multiple t-SNE runs and use evidence accumulation to determine sets of elements that are stably clustered together. We build upon per-cluster multi-parameter distribution functions to quantify cluster similarity, and merge clusters in different ensemble members. By applying the proposed approach to a time-varying ensemble, the temporal development of clusters, and in particular their stability over time can be analyzed. We apply this approach to analyze a time-varying ensemble of 3D cloud simulations. The visualizations via t-SNE, parallel coordinate plots and scatter plot matrices show dependencies between the simulation results and the simulation parameters used to generate the ensemble, and they provide insight into the temporal ensemble variability regarding the major trends in the multi-parameter distributions.