Sofia Chehab
Pronouns: she/her
Research Mentor(s): James Penner-Hahn
Research Mentor School/College/Department: Chemistry and Biophysics / LSA
Program: UROPF
Session: Session 3 (11:00am – 11:50am)
Authors: Sofia Chehab, Kevin Moser, James Penner-Hahn
Abstract
While it is possible to gain structural information about a molecule from an X-ray spectrum using an available program, Finite Difference Method Near Edge Structure (FDMNES), this process is computationally expensive to optimize. To avoid this difficulty, our group had previously developed a machine learning program that is designed to output a molecular structure by fitting an X-ray spectrum. It first takes in data about the molecule’s normal modes, a variable that describes the vibrational motions of the compound. The program then fits the X-ray spectrum and outputs the corresponding molecular structure. This process, however, needed to be refined so that it could make a more accurate fit for the spectra. One way of going about this was to increase the number of normal modes it could utilize, which would give the program more information to work with. This would allow the program to more accurately estimate a molecular structure from a computed spectrum. This is valuable for a number of reasons. For one, we might want to see how a molecule has changed shape just by looking at how its spectrum changes. Usually, this is done by hand, but by using our approach, we would not have to make any assumptions about how changes in the spectrum affect structure. In addition, our program also gives us realistic ways that the molecule could change—individual atoms rarely move without influencing the rest of the structure. Overall, our machine learning approach to obtaining structure from spectrum is faster and more consistent than fitting it by hand, and it is less computationally expensive than writing an algorithm that does an FDMNES calculation each time we change a parameter.
