INDENTIFICATION OF DESCRIPTORS IN CO OXIDATION VIA PRINCIPAL COMPONENT ANALYSIS AND PARTIAL LEAST SQUARES
Date
2019-05
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
Signi cant increases in the power and availability of computational resources
have expanded the potential of accelerating catalyst discovery via computational screen-
ing. One limitation to this endeavor is the requirement of computationally expensive
transition state calculations which are necessary to de ne a microkinetic reaction model
and determine a catalyst's activity. Linear scaling relationships provide a way to pre-
dict transition state energies from computationally cheaper adsorption energies. For a
complete reaction mechanism, the linear dependence of the energy states means that
the rate determining steps of a kinetic pro le may be adequately de ned by a few (1-2)
key descriptors. A way to automatically identify descriptors from a library of available
data is an essential step towards accelerating computational catalyst screening.
Since the cost of calculating full kinetic pro les is high, kinetic studies are often
limited to small sample sizes where the number of samples is far less than the number of
variables. Due to the existence of linear scaling in catalysis, the variable space is bound
to be highly collinear. Therefore, a method used to automatically identify catalytic
descriptors should be robust to small sample sizes and high collinearity. Statistical
methods have been developed to handle these situations. These methods are divided
into dimensionality reduction and regularized regression methods.
In this work, we used dimensionality reduction methods (principal component
analysis, partial least squares) to select descriptors. The data is primarily taken from
literature sources on CO oxidation. In these data sets, the active metal is varied across
samples, and when applicable, the support is held constant. We created models for
the activation energies using candidate descriptors of energy changes (e.g., adsorption
energies, enthalpy of elementary steps) for both extended metal surfaces, atomic-scale
metal clusters, and supported single-atom catalysts.
Dimensionality reduction methods are shown to be an excellent visualization
tool for high dimensional catalytic data. We conclude that partial least squares (PLS)
can automatically identify descriptor models in the CO oxidation datasets across a
wide range of ideal and nonideal catalysts. In addition, variable selection methods
using the PLS projections, namely variable importance in projections (VIP) and sig-
ni cance multivariate correlation (SMC), can prune irrelevant descriptors. Complete
PLS models and PLS models with pruned descriptors are benchmarked to descriptor
models identi ed in the literature, and we found PLS-SMC created the most models.
Description
Keywords
Chemical engineering,Descriptors in CO-OXIDATION