Bayesian cluster expansion with lattice parameter dependence for studying surface alloys
Embargo until
Date
2016-05-06
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Johns Hopkins University
Abstract
DFT+Cluster Expansion+Monte Carlo (DFT-CE-MC) process of studying thermodynamic properties of lattice structure is of great use in materials design for a variety of applications. Among these applications, rational design of alloy surface for catalyst usage is a promising field with great importance. Unlike the traditional experimental approach, the DFT-CE-MC process provides a foundation for a “virtual laboratory” in which all surface structures can be studies computationally with little aids of experimental data. This kind of “virtual laboratory” will save a lot of effort in experimental tries and thus speed up the design of high catalyst structures.
In details, thermodynamic stable ground states of surface alloy under different conditions and the relationship between atomic structure, formation enthalpy, chemical potential and catalyst properties can be gathered through this process. With these information, structures optimizing catalyst properties could be proposed and a detailed analyze of influential factors of catalyst properties is available.
However, a problem is raised if considering the composition mismatch between surface and bulk region. Since the lattice of surface region should always match that of bulk region, the lattice parameter of the surface, and hence the interactions among near-surface atoms, varies with the composition of the underlying bulk material.
Collecting the training data of surface structures with varies underlying bulk composition through DFT demands a large amount of calculation, and should be wasteful when the strain energy relationship can be predicted accurately through known model, such as quadratic rule.
Here, in this thesis, we demonstrate that quadratic rule model can be included in cluster expansion framework, especially for Bayesian cluster expansion approach, so that surfaces under a variety of strains can be used to train a single cluster expansion that predicts properties as a function of atomic order and surface strain. We implemented this idea on Bayesian cluster expansion and tested our method in Au-Pt and Ni-Pt alloy surface with oxygen adsorption. The result shows the validity of the idea and suggests a great potential of this method in application of catalyst research.
Description
Keywords
cluster expansion, Bayesian cluster expansion, computational materials, machine learning, catalyst