Machine Learning-Driven Chemical System Predictions and Microscopy Image Analysis For Energy Applications
Date Issued
May 2026
Author(s)
Furth, Nicholas
The University of Tennessee
Advisor(s)
Thomas Zawodzinski
Additional Advisor(s)
Sergei V Kalinin
Adam E Imel
Manolis Doxastakis
Abstract
This dissertation explores the use of machine learning (ML) models to accelerate chemical and materials systems research, focusing on two distinct areas, predicting molecular properties of microemulsions (µEs) using structural and feature-based representations, and analyzing scanning transmission electron microscopy (STEM) images to detect nanoparticles and their distributions in electrodes. Together, these projects enable high-throughput data-driven analysis of chemical and material systems leading to new possibilities in energy applications. The first section of this work investigates ML models for predicting key properties of µEs as alternative electrolytes for redox flow batteries. Graph neural networks (GNNs), along with traditional ML models including gradient boosted trees and Gaussian processes, were evaluated across property prediction tasks. The models predicted the equivalent alkane carbon number (EACN), which governs oil hydrophobicity and represents the oil in hydrophilic-lipophilic difference equation using a limited dataset, with performance compared across geometry optimizations. For critical micelle concentration, at which various electrochemical properties shift. Augmented GNNs combining graphs with non-graph features, demonstrated improved accuracy. To maximize energy density, the oil phase in μEs was designed to remain liquid at room temperature without solvent dilution. Therefore, the melting point (Tm) was constrained to best predict room temperature liquid oils, Coulomb matrices (CMs) representing the intermolecular forces that drive melting processes. In cases where the oil is not liquid at room temperature, solubility predictions were made to minimize the dilution of the oil; maximizing energy density. Variational autoencoders (VAEs) we used as graph encoders representing both the solute and solvents separately allowing for multi system predictions. The second section focuses on STEM image analysis using ML. A simulated STEM dataset was generated with unbiased segmentation to train a mask R-CNN instance segmentation model capable of identifying, localizing, and outlining nanoparticles, even in cases of particle overlap where conventional image analysis methods fail. The mask R-CNN model successfully predicted per-instance segmentation masks for nanoparticles. This dissertation contributes novel techniques and insights into how ML can be adapted to more effectively predict chemical properties and interpret microscopy data, expanding the scope of ML applications in chemical and energy research, and enabling high-throughput material and molecular discovery.
Degree
Doctor of Philosophy
Major
Data Science and Engineering
File(s)![Thumbnail Image]()
![Thumbnail Image]()
Name
FURTH-PRIMARY-2026.pdf
Size
5.73 MB
Format
Adobe PDF
Checksum (MD5)
9cdaa3bd282671c5fad44b490219b3a0
Name
1771598124716-Publication_License_Feb-02-2026_(1-2).pdf
Size
447.55 KB
Format
Adobe PDF
Checksum (MD5)
ecba28e807a90837cdf4f2baef7ed4be