Predicting Electrochemical Potential of Quinones Using Machine Learning Methods

Cetin, Elif

4. term

Chemical Engineering, Master

2025

2025-06-01

73 pages

With growing energy demands and the transition to renewable sources, the need for grid-scale energy storage systems is increasing. Quinones, a group of redox-active organic compounds that can be derived from fungi and bacteria, are promising biomolecules for use in redox flow batteries due to their tunability. This study investigates the prediction of quinone standard reduction potentials using machine learning, comparing transformer-based large language models (LLMs) and graphical neural networks (GNNs). The best-performing configurations of LLM and GNN models achieved average test set R² values of 0.734 and 0.721, respectively. However, LLMs have exhibited poorer performance on validation sets compared to test sets, indicating issues with model fitting. Within the optimal configurations, the top individual LLM and GNN models achieved an R² of 0.777 and 0.774 on the test set, respectively. While LLMs demonstrated slightly better accuracy, they require significantly higher training times and computational costs.

Machine learning ; graphical neural networks ; large language models ; quinone ; reduction potential ; ChemBERTa ; redox-flow batteries ; SMILES ; InChI

Download
View record in AAU Student Projects