Nina Andrejevic, a doctoral candidate, combines machine learning and spectroscopy to identify new and valuable properties of matter.
As a child, Nina Andrejevic was born into an architectural family and loved drawing her Serbian home and other buildings. This passion was shared by her twin sister, as well as a love for science and math. These interests converged over time to form a scholarly path, which shares some characteristics with the family profession, according To Andrejevic (a doctoral candidate at MIT in materials science and engineering).
She says architecture is both creative and technical. You try to maximize features that are important for certain functionality. Andrejevic’s machine-learning work is similar to that of architects. She believes that “we start from an empty site — which is a mathematical model with random parameters — and that our goal is to train this model, the neural net to achieve the functionality that we want.”
Andrejevic, a doctoral adviser of Mingda LI, is an assistant professor in the Department of Nuclear Science and Engineering. She is a research assistant in Li’s Quantum Measurement Group and trains her machine-learning models to find new and useful material characteristics. Her research with the lab was published in Nature Communications and Advanced Science.
Her group is particularly interested in the study of topological materials. She says these materials are an exotic form of matter, capable of transporting electrons on the surface with no energy loss. This makes them very interested in developing more efficient technologies.
Andrejevic and her sister Jovana (a doctoral candidate at Harvard University in applied physics) have created a faster and more flexible method to test material samples for topological characteristics.
She says that if the ultimate goal is to “produce better-performing, more energy-saving technologies,” then “we must first determine which materials are suitable candidates for these applications. Our research can confirm this.”
Partnering
This research began more than a decade ago. Andrejevic says, “My sister and me always thought it would be fun to do a joint project, and when Mingda suggested the study of topological material, it occurred to us that we could make it a formal collaboration.” She notes that the sisters share more academic interests than other twins. “Being a twin is a huge part of my life, and we work together well, helping each other in areas we don’t understand.” Andrejevic’s dissertation work, which encompasses several projects, uses specialized spectroscopic techniques and data analysis bolstered by machine learning, which can find patterns in vast amounts of data more efficiently than even the most high-throughput computers.
All my projects share a common thread: the desire to improve or accelerate our understanding of these characterization tools and, thereby, obtain more useful information with traditional or approximate models. This is an example of the twins’ topological material research.
Researchers must examine materials at quantum and atomic scales to discover new and useful properties. Photon and neutron spectroscopic techniques are useful in capturing previously unknown structures and dynamics. They also help to determine how heat, magnetic or electric fields and mechanical stress affect materials at Lilliputian levels. This realm is where materials behave differently than they would at the macro-scale. These laws are called quantum mechanics.
The current experimental methods for identifying topological material are difficult and inaccurate, which could lead to the exclusion of viable candidates. These pitfalls could be avoided using a widely used imaging technique called X-ray Absorption Spectroscopy (XAS) and paired with a trained neural system. XAS uses focused X-ray beams to penetrate matter and map its electron structure. It provides unique radiation data that is specific to the sampled material.
Andrejevic says, “We wanted to create a neural network that could recognize topology from a material’s XAS signature. This is a much more accessible measure than other approaches.” This would allow us to screen for a wider range of topological materials.
The researchers used two databases to feed their neural network information over months: one had topologically predicted materials, while the other had X-ray absorption data covering a wide range of materials. Andrejevic explains that the model can read new XAS signatures and determine if the material producing the spectrum is topological if properly trained.
Andrejevic says that the research team’s method has shown promising results. They have published their preprint in “Machine Learning spectral indicators for topology.”
Moving toward materials studies
Andrejevic’s first year at Cornell University was when she first felt the joy of looking at matter at an intimate level. After taking a course on nanoscience and nanoengineering, she joined a research team that studies materials at the atomic level. She says, “I feel that I am a visual person. This idea of being able see things that were up until that point just equations or concepts — this was really exciting.” “This experience brought me closer to materials science.”
Andrejevic’s doctoral research in machine learning was pivotal. She will continue to be a central figure after MIT. She will graduate this winter and head straight to Argonne National Lab, where she won the prestigious Maria Goeppert Maier Fellowship. This fellowship is awarded to “internationally outstanding doctoral scientists or engineers at early stages in promising careers.”
It will be difficult to say goodbye to her sister, whom she has never been apart from for so long. Andrejevic says, “It’ll be very different.” She adds that she hopes Jovana will continue collaborating with her, regardless of the distance.
