Machine learning interatomic potentials for first-principles electrochemistry

Loading...
Thumbnail Image
Authors
Fazel, Kamron
Issue Date
2024-05
Type
Electronic thesis
Thesis
Language
en_US
Keywords
Electrical engineering
Research Projects
Organizational Units
Journal Issue
Alternative Title
Abstract
Computational predictions of material properties are growing in reliance on machine learning tools to unlocking the calculations at the timescales and sizes to compute dynamic properties (e.g., melting temperature, solvation, diffusion, reaction rates).Machine learning methods have mostly been trained and used on the homogeneous materials with a high degree of accuracy, but have not been readily applied to mixed systems of solutes, solvents, and interfaces. Our research explores use of machine learned interatomic potentials (MLP) on complex systems (molten salts, water, and polymer catalysts) to discover first-principles based properties. First, we study homogeneous molten salts and apply techniques in MLPs to compute phase diagrams. We apply techniques to predict uncertainties of data and MLPs as they project through these calculations to find an optimal density functional theory exchange correlation functional. Then, we develop and characterize MLP performance in inhomogeneous water and molten salt critical to developing solvation models.Inhomogeneous and heterogeneous materials push MLPs to their limit and beyond, requiring new methods for developing custom training data and quality validation. We show how standard homogeneous potentials fail to capture these environments and the need to design custom training data to teach the potentials properly. We also develop an alternate physically-informed delta MLP that can similarly capture the inhomogeneous environment of surface tension. Lastly, we study a heterogeneous polymer catalyst system (in fuel cells) to characterize structure and dynamic properties.With the most complex of interfaces, we explore advanced machine learning techniques and new MLPs. These include on-the-fly learning, delta learning, and state-of-the-art message passing neural networks with equivariance to build a stable and accurate MLP. This enabled us to determine proton diffusion differences between bulk and Pt surface, surface reactions, and early polymer structuring.
Description
May2024
School of Engineering
Full Citation
Publisher
Rensselaer Polytechnic Institute, Troy, NY
Terms of Use
Journal
Volume
Issue
PubMed ID
DOI
ISSN
EISSN
Collections