Investigation of the capabilities and limitations of the Open Catalyst pre-trained machine learning models

Mulder, Goos (2024) Investigation of the capabilities and limitations of the Open Catalyst pre-trained machine learning models. Bachelor's Thesis, Applied Physics.

Preview	Text bAppP2024MulderGF.pdf Download (8MB) | Preview
	Text toestemming Goos Mulder _ degree programme_ Applied Physics.pdf Restricted to Registered users only Download (133kB)

Abstract

Adsorption energy is of great importance in material science. It refers to the energy released when a molecule adheres to a surface. Obtaining a system’s adsorption energy can help understand the underlying mechanism of the system’s macroscopic behaviour. Density Functional Theory is a widely used computational method for atomic and molecular simulations, including finding adsorption energies. DFT is an extremely computationally heavy iterative quantum mechanical simulation method. New approaches are being researched to avoid this computationally heavy method, including creating atomic force fields based on graph neural networks. The Open Catalyst Project is a collaborative project that aims to accelerates research into machine learning-based molecular dynamics models to assist or replace DFT. The project provides large datasets with over 1.2 million DFT relaxations and various pre-trained graph neural networks for the community to improve. This report contains an analysis of an invariant and equivariant model. Both models perform a set of relaxations on 19 bimetallic compounds with ammonia-base adsorbates. The mean absolute errors on these configurations for both models are found to be 1.0669 eV and 1.1066 eV for EquiformerV2 and GemNet-OC-Large, respectively. Additionally, relaxations of the more complex molecule, zinc dialkyl dithiophosphate (ZDDP), on an iron surface are performed, leading to incorrect outcomes. The models are found to be much faster when compared to DFT.

Item Type:	Thesis (Bachelor's Thesis)
Supervisor name:	Giuntoli, A.
Degree programme:	Applied Physics
Thesis type:	Bachelor's Thesis
Language:	English
Date Deposited:	03 Jul 2024 14:05
Last Modified:	03 Jul 2024 14:05
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/33016

Actions (login required)

View Item