Yuexia (pronounced “Yusia”) Lin took an unusual route to Harvard University and the Department of Energy Computational Science Graduate Fellowship (DOE CSGF).
Lin, who uses the name Luna (the first character of her given name stands for moon), was 18 when she moved from China to the United States with her mother and ultimately settled in New York City. Although she had finished high school, Lin had to repeat her education, attending night classes while working a low-wage day job.
She was interested in science, but Lin aspired to be a journalist and filmmaker. While on an internship at public radio station WNYC before college, she asked talk-show host Brian Lehrer what students should study to prepare for a journalism career.
“He said journalism, good writing, is a craft you learn by doing,” Lin recalls. Lehrer said young people should study something that would be difficult to master outside academia, such as a scientific discipline. Lin decided to follow her love of physics, first to Barnard College, and perhaps return to journalism later.
Lin deferred entry for a year after Harvard accepted her for graduate school in applied mathematics, but joined her future advisor, Chris Rycroft, on a summer project at Lawrence Berkeley National Laboratory (LBNL), where he is a visiting faculty scientist.
Their goal was to explore algorithms for diffusion-limited dissolution, a randomly determined process such as rock eroding in wind or water. The first goal was to boost Lin’s coding skills. The second was to study the properties of these dissolving shapes to find interesting geometries or other characteristics.
Since formally joining Rycroft at Harvard, Lin has mostly focused on methods to model fluid-structure interaction (FSI) – a common phenomenon in a world where water and air constantly move around objects. In the problems Lin studies, however, the structures deform in response to forces.
“We’re going into more challenging problems that might have separation of scales,” such as vats of fluid containing tiny structures like DNA or nanotubes, Lin says. Without algorithms that handle these huge spans and run on high-performance computing (HPC) systems, “these problems will be really, really hard.”
Traditional fluid mechanics divides the region to be modeled with a grid of fixed points. At each point, equations calculate the velocity and other properties of fluid moving past. Solid mechanics generally uses a moving grid, attached to the surface of the modeled object. When the solid deforms, “you’re moving with it and calculating how much stretch and compression there is,” Lin says. Connecting these perspectives for FSI complicates matters and adds computational steps.
Rycroft and colleague Ken Kamrin of the Massachusetts Institute of Technology call their approach the reference map technique. In essence, the researchers flipped the perspective, calculating quantities such as deformation and stress in the same fixed grid. “Now the computational mesh for fluids and solids are united. We no longer have to do this bridging,” Lin says. She’s tuning the method to run well on HPC platforms.
The reference map technique could make FSI problems accessible to HPC tools that accelerate calculations or make them more accurate. During her 2018 practicum with Ann Almgren in LBNL’s Center for Computational Sciences and Engineering, Lin focused on learning one: adaptive mesh refinement (AMR) as implemented in the AMReX framework. AMR uses a hierarchical structure, applying a finer grid to capture critical details where needed and a coarser one in areas with less activity, conserving computational resources.
Lin planned to learn AMReX and apply it to real FSI problems. The task proved more difficult than expected, and she left with a working but incomplete, version of her desired software. Lin returned for second practicum in 2019 to refine the code and use it to model the bloodstream movements of parasites called trypanosomes that cause sleeping sickness. Although blood is viscous, these single-cell organisms are fast swimmers. Their speed helps remove antibodies that would mark them for removal. “They swim under the radar until they reach a critical mass in our body,” causing illness, Lin says.
Lin used the Cori supercomputer at LBNL’s National Energy Research Scientific Computing Center, for her 2019 practicum. She’s also used Oak Ridge National Laboratory’s Summit.
The practicums gave her insights into HPC, Lin says, and “made me more ambitious for what kind of science I can do with this method. We can go for large-scale simulations with very tough multiscale problems” in three dimensions.
Lin expects to graduate in 2021 and seek a postdoctoral appointment. After that, she’s open to positions in university and national laboratories to further her research.
Image caption: These images from a fluid-structure interaction animation show two spheres and two cubes, with one of each pair pre-stretched and the other undeformed, as they settle under gravity, with the animation proceeding left to right and top to bottom. The colored particles are passive tracers that visualize fluid flow. The bottom wall exerts a repulsive force that stops objects from getting too close or moving through it. The first image (upper left) is the initial condition of the animation. The remaining frames show the spheres contacting the bottom wall and bouncing back, the cubes nearing the spheres and deforming due to contact, and the cubes falling between the spheres and pushing them apart. The model shows that the objects are reasonably presented even at a coarse computational resolution. Credit: Y. Luna Lin.