Alumni Projects Chosen for First U.S. Exascale Computer

Department of Energy Computational Science Graduate Fellowship (DOE CSGF) alumni are involved in at least three of 10 projects chosen to run first on the department’s next major high-performance computing (HPC) system.

The Aurora supercomputer is the first U.S. HPC system designed to achieve exascale performance – a quintillion (10¹⁸) calculations per second. It’s expected to launch in 2021 at the Argonne Leadership Computing Facility (ALCF), a DOE user facility at Argonne National Laboratory.

ALCF leaders chose 10 projects for its Early Science Program (ESP), which is designed to prepare key applications, software libraries and programming infrastructure for the new, massively parallel machine, an Argonne release says. The projects focus on handling and analyzing large amounts of data and on machine learning, in which algorithms examine new data and categorize it based on characteristics derived from previous data.

Amanda Randles, a fellow from 2010 to 2013 and now a Duke University biomedical engineering professor, is principal investigator on one ESP project, “Extreme-Scale In-Situ Visualization and Analysis of Fluid-Structure-Interaction Simulations.” The project advances data science to help understand what influences tumor cell movement in the circulatory system. Randles has previously gained attention for her models of the human vascular system.

Nicholas Frontiere, a University of Chicago student and fellow from 2013 to 2017, is part of a team working on the “Dark Sky Mining” project with Argonne’s Salman Habib. An Argonne announcement says the endeavor will connect some of the world’s biggest, most detailed cosmological simulations with data gathered by the Large Synoptic Survey Telescope, a comprehensive observation of the visible sky. Frontiere has worked to add new physics and higher precision to the simulations and to make them run well on machines that couple graphics processing units with standard central processing units. The Argonne group’s cosmic simulations also were chosen for an earlier set of Aurora ESP projects.

Julian Kates-Harbeck, a fellow who recently finished the DOE CSGF, is chief architect of an Aurora ESP code that uses machine learning to predict damaging disruptions in reactors that will use nuclear fusion, the process powering the stars, to produce energy. He started the project while on a DOE CSGF practicum with the project’s principal investigator, William Tang, at the Princeton Plasma Physics Laboratory.

With funding from two DOE agencies – the Office of Science and the National Nuclear Security Administration – the fellowship supports doctoral students in fields that use HPC to solve complex science and engineering problems. More 450 fellows and alumni are part of a community that addresses national priorities at laboratories, in academia and in industry.

Image caption: The evolution of dark matter distribution over time, from a redshift of 4 (about 12 billion light years) to today over just a piece of a simulation, about 81 million parsecs by 81 million parsecs by 41 million parsecs (around 264 million light years by 264 million light years by 134 million light years). It shows the detail the model was able to resolve in the dark matter web. Credit: Katrin Heitmann, Nicholas Frontiere, Chris Sewell, Salman Habib, Adrian Pope, Hal Finkel, Silvio Rizzi, Joe Insley, Suman Bhattacharya. The Q Continuum simulation: Harnessing the power of GPU accelerated supercomputers. The Astrophysical Journal Supplement Series, 2015; 219 (2): 34 DOI: 10.1088/0067-0049/219/2/34.