Kyle Felker was a University of Chicago physics and mathematics student when he caught the high-performance computing (HPC) bug.
The awakening began the summer after his freshman year, when he worked with a computer simulation of granular mechanics at the Universidad de Chile. “I was woefully unprepared for it. I didn’t know a single programming language,” recalls Felker, now a Department of Energy Computational Science Graduate Fellowship (DOE CSGF) recipient. Yet it exposed him to computational science and led to a master’s-level course with Andrew Siegel, a Chicago adjunct professor who models nuclear reactors at nearby Argonne National Laboratory.
“That’s what changed me, when I took that course” and learned how computational science is applied. Felker later interned with Siegel and worked with some of the country’s largest computer systems.
Felker’s interests in HPC and applied mathematics drives his research – “the reverse of what most people in the DOE CSGF do, where they have an application they’ve been interested in,” apply HPC, and “become that hybrid scientist. I was always interested in HPC first and searching for an application.” An offer from Princeton University’s James Stone to devise models of black hole mergers and other massive events tapped Felker’s fascination with astrophysics.
It’s difficult to accurately calculate the physics of these giant events, particularly their magnetohydrodynamics (MHD) – how radiating, magnetically charged gases flow. Stone has developed such simulations by building Athena, a grid-based astrophysics code. Computers calculate interactions at points on the grid. Taken together, they portray the larger phenomena, similar to how pixels comprise a digital photo.
Historically, such codes have been first- or second-order accurate: straightforward to run but with a certain amount of inherent error. Felker is reaching for as much as fourth-order accuracy, reducing error as the calculations converge on an answer.
“I’m trying to make a practical tool for astrophysicists that takes a lot of the elegant, complex numerical methods” developed at Lawrence Berkeley National Laboratory and other institutions, “and turn them into a technique that is more computationally intensive per step, but you’re going to get more accurate solutions in the long run.” The techniques aren’t commonly used in astrophysics, Felker says, but should run well on new HPC architectures, including those using manycore chips and a combination of processor types.
Felker has implemented his fourth-order accurate method in an unreleased section of Athena. Now he’s preparing to test it on astrophysical problems using HPC systems.
To that end, Felker spent part of summer 2017 at the Argonne Training Program on Extreme-Scale Computing, learning the skills necessary to run on leading HPC systems. “I came into grad school with a lot of HPC experience, but I’ve been on the method side for the last few years and things change a lot in HPC,” Felker explains. He’s also obtained an allocation on Cori, a Cray XC40 based on Intel manycore chips, at the National Energy Research Scientific Computing Center.
Felker didn’t go far for his 2016 summer practicum, but he says the project he worked on at the Princeton Plasma Physics Laboratory (PPPL) differed the most from his thesis research of any he considered. With scientist William Tang, Felker helped improve a machine-learning algorithm to predict plasma disruptions in tokamaks. Physicists use these donut-shaped fusion energy reactors to heat hydrogen isotopes to temperatures as hot as the sun, fusing their nuclei to release tremendous energy. But the plasma can escape the magnetic field containing it, damaging the tokamak walls.
The algorithm analyzes data from previous fusion experiments, learning the signs preceding disruptions so it can monitor incoming sensor data and predict when one is imminent.
The practicum was unusual because Felker worked with Harvard University’s Julian Kates-Harbeck, another DOE CSGF recipient also doing a PPPL practicum. Kates-Harbeck had experience with machine learning and had started to tune the code when Felker arrived. Felker had never worked with the technique, so having an expert to collaborate with was a bonus.
Although not a plasma physicist, Felker contributed knowledge gained from modeling astrophysical plasmas. The fellows were able to test the algorithm against alternative tools and gave a talk about it at General Atomics, a fusion research contractor. There are plans to publish a paper.
Felker plans to graduate by fall 2018 and hopes to work at a national laboratory or in industry. By then he may have a more succinct way to describe his interests. While he studies astrophysics, “I always preface that with ‘I’m not actually an astrophysicist’” but more a computational scientist or applied mathematician. “I’m still searching for that precise definition.”
Image caption: The density evolved by Kyle Felker’s fourth-order algorithm in a multidimensional magnetohydrodynamics vortex problem widely used to test a code’s ability to robustly capture turbulence and shock interactions. The X and Y axes indicate position. The colors correspond to gas density, with blue/purple indicating lower density and yellow/red indicating high density. Image courtesy of Kyle Felker.