Project Overview

Machine Learning for Broadband Rotational Spectroscopy

Faculty Sponsor

Dan Zaleski (dzaleski@colgate.edu)

Department(s)

Chemistry
Computer Science

Abstract

Machine learning, commonly referred to as artificial intelligence, is a powerful tool for identifying patterns and solving permutable problems. Artificial neural networks (ANN) – named because they operate in an interconnected fashion similar to our brains – offer chemistry a major opportunity because they provide one way in which machines are able to learn and make determinations about data. ANNs are able to understand complex systems and achieve difficult goals, and in some cases, better than humans.
 
Since the invention of chirped-pulse Fourier transform microwave (CP-FTMW) spectroscopy, the increase in throughput and spectral complexity has outpaced the ability to assign and analyze acquired data. Fortunately, high-resolution broadband rotational spectroscopy seems to be particularly suitable for machine learning. Rich and repeating patterns formed by molecular rotational transitions can be observed over ∼10 GHz of spectral breadth in a typical chirped-pulse experiment. At the same time, fine linewidths of ∼1 MHz and unsurpassed sensitivity of transition frequencies to molecular geometry enable precise and unambiguous determination of molecular parameters. The goal of this work to leverage machine learning for the identification and assignment of broadband rotational spectra. The hope is to make broadband rotational spectroscopy as fully automated as possible. If spectra can be accurately assigned by a machine that can learn, the whole process becomes much more portable and accessible since the amount of technical expertise needed is reduced.
 
During this project, students will have the opportunity to develop two different types of neural networks. The first is a basic feed forward neural network for classification and “fitting” of rotational spectra with complex features (e.g. nuclear hyperfine structure). The second is a long short-term memory (LSTM) neural network for deconvolving (or separating) complex mixtures. Greedy algorithms and genetic algorithms may also be explored. GPU computing will be emphasized.
 
Students are not required to have prior coding experience, but they will be expected to develop the skill over the course of the project. Coursework in physical chemistry, physics, or computer science is preferred, but not strictly required. Students with at least an interest in physical chemistry are especially encouraged to apply.

Student Qualifications

Students are not required to have prior coding experience, but they will be expected to develop the skill over the course of the project. Coursework in physical chemistry, physics, or computer science is preferred, but not strictly required. Students with at least an interest in physical chemistry are especially encouraged to apply.

Number of Student Researchers

2 students

Project Length

10 weeks


Applications open on 01/03/2020 and close on 02/05/2020


<< Back to List





If you have questions, please contact Karyn Belanger (kgbelanger@colgate.edu).