It’s common to hear astronomers say that time spent star-gazing as a child helped inspire them.
Harshil Kamdar isn’t one – although he earned bachelor’s degrees in astronomy and physics at the University of Illinois at Urbana-Champaign and now is a Harvard University doctoral student in computational astrophysics. “I never, when I was 7 or 8, used to look at the night sky and wonder,” says Kamdar, a Department of Energy Computational Science Graduate Fellowship (DOE CSGF) recipient.
“There never was one defining factor that led me to pursue astronomy. It was a lot of small things,” especially studying with astronomers Robert Brunner and Matthew Turk. They drew Kamdar into research that incorporated computer science, another of his interests.
Now Kamdar models how stars form and move through our galaxy. His research capitalizes on data from Gaia, a satellite that tracked the positions and velocities of a billion stars. Kamdar and his advisor, Charlie Conroy, interpret that information to help understand how the Milky Way evolved.
It’s common for computational astronomers to simulate galaxies, but because of limited computer power and the system’s immense scale, their models clump stars into particles, each about a thousand times the sun’s mass, to make them feasible.
Milky Way star clusters, however, are about 500 solar masses big, Kamdar says. Earlier simulations “are not resolving the clustered nature of star formation and the things we can study from Gaia.” The code he and his colleagues are developing portrays each star individually. “We let billions of these evolve in our simulations and then try to make mock star catalogs to compare to what Gaia sees.”
To make evolving billions of star-sized particles computationally feasible, the researchers approximate their gravitational interactions. Instead of precisely calculating those forces, they give the system a nudge based on estimates of how much force the stars would feel.
The team’s simulations follow as many as 4 billion stars as they form in clusters and dissipate through the galaxy, like a drop of ink spreading as smaller dots through water. Stars that move in similar ways and at similar speeds, according to Gaia, probably were born in the same cluster. At the same time, ground-based spectroscopy shows that stars with similar chemical makeup also are likely to have a common birthplace.
In a paper published on line in March 2019, Kamdar and his colleagues used their model to demonstrate that Gaia data and chemical composition information combined can better identify stars that were born together than either alone. In a second paper posted a month later, Kamdar and several colleagues showed that 111 co-moving pairs – star couples traveling in similar ways, even though they are far apart – probably were created in the same cluster. The researchers supported that conclusion by showing the co-moving pairs each had similar chemical compositions.
By comparing their models with real data, Kamdar and his fellow researchers identify weaknesses in them and learn about physics governing star formation and the galaxy. In the end, they hope to characterize how galaxies like ours are born and evolve.
Most of Kamdar’s calculations ran on Odyssey, Harvard’s computing cluster, but he also used Blue Waters, a National Science Foundation supercomputer based at the University of Illinois. He gained access to that machine as an undergraduate, when he was chosen as a Blue Waters Fellow. He developed a machine-learning algorithm that identified locations in dark matter-only models where galaxies were likely to form. Dark matter’s gravity influences how visible matter clumps together.
Kamdar also used machine learning during his 2018 practicum with data scientist Caleb Phillips at the National Renewable Energy Laboratory in Colorado. They worked to identify crystals with favorable band gaps, the quality that governs their efficiency in converting sunlight to electricity.
Crystal band gap calculations using standard techniques are demanding. To cut development time, Kamdar trained a neural network, a kind of machine learning algorithm, on these data and used it to predict band gaps for crystals based only on their structures. The algorithm performed reasonably well, Kamdar says, but wasn’t accurate enough to be useful. A paper on the project is in the works.
Kamdar expects to graduate in 2021 and then hopes to enter academia. In the meantime, he’ll also continue sharpening his science communication skills, an interest he developed by participating in ComSciCon, an annual gathering to help graduate students improve their ability to relate their research to nontechnical audiences. “It’s been a really big part of my graduate career in the past year or two, and I definitely want to continue on that path.”
Video caption: A simulation of our galaxy, the Milky Way, with 4 billion stars evolved over the last 5 billion years. The colors correspond to stars that are born in different clusters. This simulation is the first of its kind that can resolve individual stars and, consequently, allows us to pursue galactic archaeology – the study of understanding the galaxy’s distant past using the stars we see today – in unprecedented detail. Credit: Harshil Kamdar.