Doctoral candidate Nina Andrejevic combines spectroscopy and machine studying strategies to establish novel and priceless properties in matter.
Born right into a household of architects, Nina Andrejevic beloved creating drawings of her house and different buildings whereas a baby in Serbia. She and her twin sister shared this ardour, together with an urge for food for math and science. Over time, these pursuits converged right into a scholarly path that shares some attributes with the household career, in response to Andrejevic, a doctoral candidate in supplies science and engineering at MIT.
“Structure is each a inventive and technical subject, the place you attempt to optimize options you need for sure sorts of performance, like the scale of a constructing, or the format of various rooms in a house,” she says. Andrejevic’s work in machine studying resembles that of architects, she believes: “We begin from an empty website — a mathematical mannequin that has random parameters — and our purpose is to coach this mannequin, referred to as a neural community, to have the performance we want.”
Andrejevic is a doctoral advisee of Mingda Li, an assistant professor within the Division of Nuclear Science and Engineering. As a analysis assistant in Li’s Quantum Measurement Group, she is coaching her machine-learning fashions to hunt for brand new and helpful traits in supplies. Her work with the lab has landed in such main journals as Nature Communications, Superior Science, Bodily Overview Letters, and Nano Letters.
One space of particular curiosity to her group is that of topological supplies. “These supplies are an unique part of matter that may transport electrons on the floor with out vitality loss,” she says. “This makes them extremely fascinating for making extra energy-efficient applied sciences.”
Along with her sister Jovana, a doctoral candidate in utilized physics at Harvard College, Andrejevic has developed a technique for testing materials samples to foretell the presence of topological traits that’s sooner and extra versatile than different strategies.
If the final word purpose is “producing better-performing, energy-saving applied sciences,” she says, “we should first know which supplies make good candidates for these purposes, and that’s one thing our analysis may also help verify.”
The seeds for this analysis had been planted greater than a yr in the past. “My sister and I all the time stated it will be cool to do a mission collectively, and when Mingda urged this examine of topological supplies, it occurred to me that we may make this a proper collaboration,” says Andrejevic. The sisters are extra comparable than most twins, she notes, sharing many educational pursuits. “Being a twin is a large a part of my life and we work collectively effectively, serving to one another in areas we don’t perceive.”
Andrejevic’s dissertation work, which encompasses a number of tasks, makes use of specialised spectroscopic strategies and information evaluation, bolstered by machine studying, which may discover patterns in huge quantities of information extra effectively than even probably the most high-throughput computer systems.
“The unifying thread amongst all my tasks is this concept of making an attempt to speed up or enhance our understanding when making use of these characterization instruments, and to thereby receive extra helpful data than we will with extra conventional or approximate fashions,” she says. The twins’ analysis on topological supplies serves as a working example.
With a purpose to tease out novel and probably helpful properties of supplies, researchers should interrogate them on the atomic and quantum scales. Neutron and photon spectroscopic strategies may also help seize beforehand unidentified buildings and dynamics, and decide how warmth, electrical or magnetic fields, and mechanical stress have an effect on supplies on the Lilliputian stage. The legal guidelines governing this realm, the place supplies don’t behave as they could on the macro-scale, are these of quantum mechanics.
Present experimental approaches to figuring out topological supplies are difficult technically and inexact, probably excluding viable candidates. The sisters believed they may keep away from these pitfalls utilizing a broadly utilized imaging method, referred to as X-ray absorption spectroscopy (XAS), and paired with a skilled neural community. XAS sends centered X-ray beams into matter to assist map its geometry and electron construction. The radiation information it gives affords a signature distinctive to the sampled materials.
“We wished to develop a neural community that would establish topology from a fabric’s XAS signature, a way more accessible measurement than that of different approaches,” says Andrejevic. “This might hopefully enable us to display screen a much wider class of potential topological supplies.”
Over months, the researchers fed their neural community data from two databases: one contained supplies theoretically predicted to be topological, and the opposite contained X-ray absorption information for a broad vary of supplies. “When correctly skilled, the mannequin ought to function instrument the place it reads new XAS signatures it hasn’t seen earlier than, and tells in the event you if the fabric that produced the spectrum is topological,” Andrejevic explains.
The analysis duo’s method has demonstrated promising outcomes, which they’ve already revealed in a preprint, “Machine studying spectral indicators of topology.” “For me, the fun with these machine-learning tasks is seeing some underlying patterns and having the ability to perceive these when it comes to bodily portions,” says Andrejevic.
Shifting towards supplies research
It was throughout her first yr at Cornell College that Andrejevic first skilled the pleasure of peering at matter on an intimate stage. After a course in nanoscience and nanoengineering, she joined a analysis group imaging supplies on the atomic scale. “I really feel I’m a really visible individual, and this concept of having the ability to see issues that as much as that time had been simply equations or ideas — that was actually thrilling,” she says. “This expertise moved me nearer to the sector of supplies science.”
Machine studying, pivotal to Andrejevic’s doctoral work, can be central to her life after MIT. When she graduates this winter, she heads straight for Argonne Nationwide Laboratory, the place she has gained a prestigious Maria Goeppert Mayer Fellowship, awarded “internationally to excellent doctoral scientists and engineers who’re at early factors in promising careers.” “We’ll be making an attempt to design physics-informed neural networks, with a concentrate on quantum supplies,” she says.
This may imply saying goodbye to her sister, from whom she has by no means been separated for lengthy. “Will probably be very totally different,” says Andrejevic. However, she provides, “I do hope that Jovana and I’ll collaborate extra sooner or later, irrespective of the space!”