1 of 1

FTK: A Spacetime Meshing Framework for Robust and Scalable Feature Tracking

5D blob tracking in WDMApp

Parallel performance on Theta

Spacetime triangulation of 3D simplicial prism mesh (left) and 4D regular grid mesh for feature tracking

Two-pass trajectory reconstruction of 0D features in 3D (left) and 4D (right) in spacetime

Scientific Achievement

A high-dimensional simplicial meshing framework that simplifies, scales, and delivers feature tracking algorithms for scientific data

Significance and Impact

The use of simplicial spacetime meshing reduces ambiguities, simplifies handling of degeneracies, and scales performance for tracking scientific features such as critical points, vortex core lines, and levelsets.

Research Details

4D triangulation enables conforming subdivision of simplicial prism meshes and regular grid meshes in spacetime
Robust detection and tracking of critical points, vortex core lines, and levelsets enabled by 4D simplicial meshes
GPU acceleration of root-finding algorithms for localizing features in individual spacetime mesh cells
Load-balanced and asynchronous union-find enables scalable connected component labeling that tracks features in spacetime
Applications of tracking blob filaments in 5D XGC simulations across multiple poloidal planes
FTK software: https://github.com/hguo/ftk
Publications: Guo et al., IEEE TVCG, 27(8):3463-3480, 2021; Xu et al., IEEE TVCG, 27(6):2808-2820, 2021 (Best Paper Award in IEEE PacificVis 2021)

Robustness evaluation

Hanqi Guo (ANL), David Lenz (ANL), Iulian Grindeanu (ANL), and Tom Peterka (ANL)

Technical Contacts: �Hanqi Guo <hguo@anl.gov> (ANL)

�Supported by:

ECP CODAR, ANL LDRD Innovate��Notes:

In the exascale computing era, as the gap increases between the data production rates and available I/O bandwidth, in situ feature tracking makes it possible to derive insights during scientific simulations and thus reduce the data to be stored. FTK provides a generalized framework for in situ feature tracking. Feature tracking is important because all sciences need feature tracking to understand the evolution of features over time. The novelty of FTK is based on the generality of the framework and the ability to incorporate state-of-the-art deep learning, statistical, and topological feature tracking algorithms. FTK is also scalable and thus enables in situ feature tracking with the simulations that run on today’s and future leadership computing facilities. The building blocks of FTK include a tracking graph data structure, high-dimensional meshing, and graph-based feature abstractions. The tracking graph characterizes the birth, death, merge, and split events in time-varying data. High-dimensional meshing eliminates the need for redundant efforts to develop domain-specific feature tracking algorithms by generalizing given 2D/3D meshing into the 3D/4D spacetime domain. The graph-based representation is capable describing features in arbitrary dimensions as a subgraph of the high-dimensional mesh. In addition, FTK also incorporates distributed and parallel processing for in situ data analysis and visualization.