TASK 4: Research uncertainty estimation and visualization techniques for science-based energy system simulations.

colhead_uncertaintyWith large computational simulations there is substantial uncertainty inherent in any prediction of science-based systems. A number of factors contribute to uncertainty, including experimental measurements, mathematical formulation, and the way different processes are coupled together in the numerical approach for simulation. Tracking of and analysis of this uncertainly is critical to any work that will truly impact the creation of future energy systems. At present, the preferred technique for constructing scientific simulation data sets containing uncertainty is to construct initial conditions and model parameters by sampling from some predefined input space, execute a full simulation run on each of those samples, and then aggregate the results into an ensemble of runs. Such an ensemble has many values for each value at each point instead of the single value present in a deterministic data set containing results from a single simulation run. To quantify the uncertainty present in the simulation we consider these sets of values as samples from some underlying probability distribution. As such, we can bring to bear the entire body of statistical analysis methods to make sense of what the data tell us. Uncertainty visualization is challenging because we have relatively little intuition regarding effective ways to display sets of probability distributions.

We propose to extend our Ensemble-Vis framework to support the visual analysis of ensemble data with a focus on the discovery and evaluation of simulation outcomes. Our approach combines a variety of statistical visualization techniques to allow scientists to quickly identify areas of interest, ask quantitative questions about the ensemble behavior, and explore the uncertainty associated with the data. By linking scientific and information visualization techniques, Ensemble-Vis provides a cohesive view of the data that permits analysis at multiple scales from high-level abstraction to the direct display of data values. Ensemble-Vis is developed as a component-based framework so we will be able to connect to other software systems, e.g. Uintah and VisTrails.


The improved ability of visualization procedures in generating impressive representational imagery from complex scientific datasets is driving the use of virtual environments. As visualizations go beyond data presentation for promotional tasks and decision-making scenarios, it becomes increasingly important to understand and convey the existence of uncertainty. To this end, knowledge of the uncertainty that accompanies data and is incurred throughout the visualization process, is mandatory. During the second quarter of our project, newly-hired postdoctoral research fellow, Dr. Joel Daniels, worked on creating a new software framework for visualizing and assessing uncertainty called QuizLens; a tool that aids the user in gaining knowledge of the uncertainty accompanying visualized data.

QuizLens provides an interactive exploration framework in which users navigate their dataset using an uncertainty lens to affect local visualization views of the data within the global context. Our method employs a user-controlled portal that focuses on a region of interest and deploys multiple visualization lenses. The user is able to toggle between the different uncertainty lenses to choose an appropriate display method for the region of interest. This approach to uncertainty visualization addresses many challenges, including visual clutter, arbitrary emphasis on uncertain regions, and the displacement of multiple visualization techniques in separate windows.

We demonstrate the utility of the QuizLens framework through the exploration and visualization of a segmented brain ensemble. This dataset describes the probabilities of point locations belonging to each of eleven different segmentation types: background, cerebrospinal fluid, fat, muscle, skin, skull, connective tissue, lesion, as well as grey, white, and glial matters. The data is composed by accumulating the tissue classifications assigned to a point location across multiple segmentation results from a scanned brain to determine probability statistics.




Figure 1: QuizLens allows the user to browse local uncertainty information within a global context. Above, the user culls the volume rendering of the skin tissue probability in order to locally visualize the maximal boundary surface of the skull (left). This isosurface is extracted using dual contouring, and colored according to the variance of probability values in the neighborhood of each vertex. The visualization tasks of the global environment and uncertainty lens can be controlled independently of each other, displaying the connectivity tissue isosurface within the grey-matter volume render (right).