This article appeared in Catalyst Magazine, Spring 2024
Faculty Profile
To model systems that change over time and space—in fields from physics and chemistry to economics and computer science—researchers use partial differential equations. These equations are powerful tools for predicting dynamic changes. However, they are also notoriously complex and difficult to solve.
Aditi Krishnapriyan, who joined the faculty as an assistant professor in Chemical and Biomolecular Engineering in January, is developing machine learning methods that can solve partial differential equations to tackle these kinds of complicated simulations. She calls the connection between machine learning and science "a two-way street". She uses physics knowledge to improve machine learning methods, and machine learning methods to tackle challenges in physics and chemistry.
"It's an exciting area because there's so much to be done," says Krishnapriyan. "I'm building a group with expertise across both computer science and the physical sciences, so we can aim to iterate and make progress much more quickly." Before joining the College, Krishnapriyan was an Alvarez Fellow in Lawrence Berkeley National Laboratory's Computational Research Division.
Modeling Materials
At UC Santa Barbara, she started her undergraduate education intending to focus on chemistry. After two years she added a second major: physics.
While some of her classmates enjoyed the experimental aspects of physics and chemistry, Krishnapriyan was drawn to the theoretical side. She joined a research lab studying condensed matter and solid-state physics and began research using fundamental physics equations to model the properties of the materials.
"I enjoyed doing the theory work where I could start out thinking through my ideas mathematically with pencil and paper, but then be able to verify everything with computer simulations," she says.
Integrating Machine Learning
While in graduate school at Stanford, she also became particularly interested in how emerging machine learning methods could improve current scientific models. The methods at the time, however, fell short for the sort of problems she wanted to solve.
As part of her Ph.D., Krishnapriyan developed machine learning models that would capture the geometric and topological information of molecules and materials.
During graduate school, Krishnapriyan thought about working in industry, using her physics-inspired machine learning models to tackle broad scientific problems. This included working at Google and at Toyota Research Institute. She enjoyed the experiences in industry and and underscored the importance of scientists knowing how to code themselves.
Then, she was awarded the prestigious Luis W. Alvarez Fellowship in Computing Sciences to work as a postdoctoral fellow at Lawrence Berkeley National Laboratory.
An Interdisciplinary Lab
Krishnapriyan was attracted to Berkeley, she says, by the plethora of interdisciplinary research. Being among a diverse collection of researchers helps to provide her with fodder for improving her machine learning models and finding ways to apply them to new types of scientific problems.
Her lab focuses on developing new machine learning methods for scientific problems, ranging from developing neural network architectures to better training procedures and evaluations. She believes that these systematic strategies will allow machine learning to bridge the gap between models and the real world. Ultimately, she hopes her models will help scientists use machine learning in a more useful and accurate way.
"I am hoping the methods I develop will make it easier for scientists to take these new methods and get more suitable answers."