New Tools for Visual Analysis of Multivariate Data from Large Scale Climate Simulations

The POP ocean model is used to compute global velocity, temperature, salinity, and surface conditions at .1o resolution and 40 irregularly spaced depths, resulting in data sets 80 Gb in size for 200 time steps. This is too large for most screens and most programs; POPTEX, a visualization application developed at LANL uses low level graphics primitives to accelerate animation to a sufficiently high frame rate. Today's graphics cards employ increasingly powerful processors that rival if not exceed the capabilities of the CPU, although the instruction sets and data paths are highly optimized for graphics processors. However, there is now enough power and flexibility to migrate many algorithms to the card level, increasing the speed sufficiently to reach interactive levels even for the larger data sets. The data from the simulations are markedly different from those from observations; they are an attempt to model the processes that control the evolution of a dynamic system. We discuss a framework that we are developing that provides effective analysis and visualization tools to the user that has the ability to evaluate development and processes over a wide range of time scales and distances. First, the data is transformed from a Cartesian grid to a (streamline, streamfront, vorticity) framework. This generalization of particles and tracers allows the user to distinguish between transport and transformation processes. Multiple variables can be viewed concurrently by brushing or selecting regions in the joint probability distribution windows. The path of a packet or particle set in space and time can also be displayed as a trajectory in the distribution space. Finally, and most specific to the essential difference between simulation data and observed data, the trajectories can be linked back to the governing equations. Nonlinearities are detectable in the visualization by relating them to innovations, instances that cannot adequately be predicted from linear perturbations of operators which can be implemented on the graphics cards for very rapid computation and analysis.