Abstract
Numerical simulation and scientific visualization are often used by scientists to help them understand physical phenomena. One approach taken by some visualization systems is to identify and quantify coherent features in a simulation and track their trajectories as they evolve over time. Such feature-tracking systems operate either by relying on manual (human) efforts, or by utilizing ad hoc programs embodying heuristics that are computationally expensive to use. Our research demonstrates the use of inductive learning to construct feature-tracking programs for fluid flows. Our approach uses manually generated feature trajectories as training data, and applies inductive learning to construct feature-tracking rules that can then be incorporated into a feature-tracking program. This results in a more efficient system that can match up objects across large time steps without inspecting intermediate steps. We demonstrate our approach on the problem of tracking vortices in turbulent viscous fluids.