University of Notre Dame
Browse
DOCUMENT
HerrJE042020D.pdf (20.82 MB)
DOCUMENT
HerrJE042020D_Supplemental.pdf (2.04 MB)
1/0
2 files

Development of Neural Network Models for Prediction of Molecular Properties

thesis
posted on 2020-04-21, 00:00 authored by John E. Herr

Machine learning (ML) has seen renewed interest with the advent of modern computational accelerators such as GPUs and the availability of larger scale datasets. ML potential energy models are promising the accuracy of quantum methods at significantly reduced cost. High accuracy potential methods which scale well for long simulations are desirable for many research areas. This dissertation describes methods for collecting datasets of nonequilibrium geometries for training ML potential energy models and shows that enhanced sampling methods or methods which do not follow Boltzmann statistics are necessary for diversely sampling geometries. Several potential models are described. The first is a combination of an ML model with the many-body expansion for condensed phase systems. The result is a highly efficient and accurate method for running large simulations of liquid phase systems. The method is implemented for methanol, but extension to other systems or mixed non-covalent fragments is trivial. The next two chapters describe the development of TensorMol, and improvements made to the model to accommodate more elements. TensorMol is a general NN potential for small organic molecules which includes explicit long-range interactions. The Coulomb and van der Waals energies included with TensorMol are crucial for running accurate simulations with correct long-range behavior. The follow-up work improves upon TensorMol by reframing the parameterization such that it is constant with the number of unique elements in the training data. The result is a model which is able to treat a vast expansion of organic molecules by including eleven unique elements in the training data. This dissertation also covers the early stages and future directions of a project using graph neural network models to predict reactions products, yields, and stereoselectivities. Finally, a collaborative project to model the Stokes shift observed in photoluminescence experiments of lead-halide perovskites describes the electronic structure of these materials to elucidate the cause of the Stokes shift. A low-lying gap state is observed which follows the trends of size-dependence in experimental results. Electronic structure calculations verify this low-gap state is responsible for the size-dependence of the Stokes shift.

History

Date Modified

2020-05-22

Defense Date

2020-03-25

CIP Code

  • 40.0501

Research Director(s)

Paul Helquist

Committee Members

Alex Kandel Steven Corcelli

Degree

  • Doctor of Philosophy

Degree Level

  • Doctoral Dissertation

Language

  • English

Alternate Identifier

1155112205

Library Record

5503686

OCLC Number

1155112205

Program Name

  • Chemistry and Biochemistry

Usage metrics

    Dissertations

    Categories

    No categories selected

    Exports

    RefWorks
    BibTeX
    Ref. manager
    Endnote
    DataCite
    NLM
    DC