Descriptor based adsorption-energy prediction on nano-clusters for catalyst selection

Loading...
Thumbnail Image

URL

Journal Title

Journal ISSN

Volume Title

Perustieteiden korkeakoulu | Master's thesis

Date

2018-11-06

Department

Major/Subject

Physics of Advanced Materials

Mcode

SCI3057

Degree programme

Master’s Programme in Engineering Physics

Language

en

Pages

6+51

Series

Abstract

Nano-catalyst design, supplanting critical/rare metals with earth-abundant elements, for hydrogen evolution reactions (HERs) is a significant material-science and economic challenge. These design challenges can be significantly overcome by extensive first principle simulations. However, these simulations are computationally costly. With an introduction of descriptor, machine-learning methods afford significant advantages in such scenario. Descriptors: Smooth Overlap of Atomic Positions (SOAP) based on charge density, BLEACH, and Local Many- Body Tensor Representation (LMBTR) are proposed, and developed. These are evaluated on database of AuCu, and MoS2 nano-clusters. A learning error of 0.05 eV in adsorption energy, and 1.7 me in charge on hydrogen are realised for LMBTR; while 0.08 eV and 13.4 me, respectively for BLEACH on the AuCu dataset. Although not as accurate as state-of-the-art SOAP-lite (3.36 meV and 0.077 me, respectively), these descriptors have their own benefits. While LMBTR allows for bi-directional operability, BLEACH provides element-agnosticism; both of which missing are from SOAP.

Description

Supervisor

Foster, Adam

Thesis advisor

Jäger, Marc

Keywords

catalyst, descriptor, energy prediction, machine learning, material science

Other note

Citation