Using Delta-Machine Learning to Accurately Predict Molecular Polarizabilites
Restricted (Penn State Only)
- Author:
- Chaudhry, Imran
- Graduate Program:
- Chemistry
- Degree:
- Master of Science
- Document Type:
- Master Thesis
- Date of Defense:
- June 26, 2023
- Committee Members:
- Lasse Jensen, Thesis Advisor/Co-Advisor
Edward Patrick O'Brien, Jr., Committee Member
Philip Bevilacqua, Program Head/Chair
Benjamin James Lear, Committee Member
Sharon Xiaolei Huang, Committee Member - Keywords:
- Raman Spectroscopy
Electronic Structure Theory
Machine Learning
Chemistry - Abstract:
- In this thesis we introduce the use of a ∆-machine learning model to accurately predict the polarizability tensor. We aim to use this model to facilitate a computationally efficient method for generated surface-enhanced Raman spectroscopy (SERS) spectra. We first cover the exploration of the orientation and electronic structure of ferrocene/ferrocenium (Fc/Fc+) systems in order to explain the findings from electrochemical-SERS experiments. We used a body-centered rotation scheme to determine the configurations that Fc and Fc+ were most likely to take on the metal surface. We found that it was most likely that Fc would be generally disordered, meaning that it would not take any particular configuration on the surface. Fc+ would most likely be ordered, meaning that it would take a specific configuration on the surface. We also found that there was a difference in intensity between the Fc and Fc+ spectra of an order of magnitude. This is due to the narrowing of the HOMO/LUMO gap in Fc+ systems and the rearrangement of the energy states such that the tail mode was the most polarized mode. We then introduce the use of ∆-machine learning (∆-ML) to predict isotropic polarizability values and the polarizability tensor. The ∆-machine learning method focuses on mapping qualitative differences in property to molecular structure, allowing for the use of computationally inexpensive calculations to yield accurate results. Overall, we found that the ∆-ML models succeeded in predicting polarizability values. These predictions were more accurate than the predictions from a traditional ML model. In order to develop the most accurate ML models, we explored the uses of different similarity metrics and molecular representations. We found that the ∆-ML model performed best with the sorted Coulomb matrix representation using a Laplacian similarity kernel. Implementation of our ∆-ML model is relatively simple and can be used for both the prediction of isotropic polarizabilities and polarizability tensors. We anticipate that this model can be extended to larger systems that include transition metals, ultimately leading to the development of a machine learning framework for Raman spectra.