Feature extraction and visualization algorithm improves functional memory and research outcomes in multiphase flow simulation analysis

By Galen Fader
Oak Ridge National Laboratory

A joint research team funded by the Department of Energy’s Exascale Computing Project (ECP) has created a novel end-to-end analytics and visualization program for solving domain-specific issues in multiphase flow simulations. The program operates alongside MFIX-Exa, a massively parallel CFD-DEM (computational fluid dynamics–discrete element model) code designed to simulate and study multiphase flows. The researchers developed a novel in-situ feature extraction and data reduction algorithm that can be run at scale along with MFIX-Exa, producing informative and compact feature-preserving data outputs. These compacted data outputs can be used to perform meaningful visual analysis of MFIX-Exa data. This work was produced as a collaboration between the ECP ALPINE project, the ECP Cinema team, and ECP MFIX-Exa developers. The researchers’ work was published in the September 2022 issue of the Journal of Computational Science.

The application of advanced computing to scientific research is creating an increasing number of high-fidelity and high-resolution computational simulations to study physical phenomena. These simulations produce very large amounts of data which is generated at a significantly faster rate than that of disk storage. Without efficient data extraction and reduction, much of the information being produced is discarded without opportunity for analysis and visualization, ultimately decreasing scientists’ rate of progress.

The researchers’ primary focus was on accurately detecting regions of low particle concentration (“bubbles”) within the raw MFIX-Exa particle data while solving the disk I/O problem. They showed that their in-situ algorithm generates highly compact outputs that preserve bubble features accurately and with high temporal fidelity while remaining small enough for persistent storage and later analysis and visualization. The researchers also developed several interactive visual analytics tools for studying three-dimensional spatiotemporal bubble dynamics using the bubble features generated by the in-situ algorithm.

The researchers’ novel in-situ feature extraction algorithm can find relevant features within the extremely large datasets generated by modern exascale supercomputers. Algorithms of this type are essential for meaningful and efficient analysis of simulation data, as these data invalidate traditional offline analyses. Furthermore, this work contributes the DOE’s energy assurance efforts by supporting MFIX-Exa in the development and large-scale deployment of commercial carbon capture technologies.

Soumya Dutta, Terece Turton, David Rogers, Jordan Musser, James Ahrens, and Ann Almgren. “In situ feature analysis for large-scale multiphase flow simulations.” 2022. Journal of Computational Science (September).