136
T. Toyao et al.
1 Introduction
Much effort has been devoted to the development of new catalysts for more efficient utilization of CH 4 owing to its economic importance and significant interest in
both industry and academia. The development of such catalysis remains challenging
because the complexity of the surface reactions and the large number of independent
parameters make the prediction of catalyst performance a formidable task. Studies
have recently applied descriptors (indicators or relations) to solid catalysts to find
a relationship between the bulk/surface structure and performance [1]. In principle,
the pace of catalyst discovery could be accelerated by using computational screening
methods [2]. Recent developments in methodology and computer technology, as
well as the establishment of a descriptor-based approach for the analysis of reaction
mechanisms and trends across the periodic table, allow fast screening of new catalytic
materials, which has already resulted in the computational discoveries of several new
materials [3, 4]. However, these computation-based approaches are typically based
on first-principles calculations and thus usually have a high computational cost.
In recent years, machine learning (ML) methods have gained increasing popularity
in the molecular and materials science communities for use in the high-throughput
screening or the prediction of various kinds of physical properties whose mathematical modeling involves complex principles [5, 6]. ML methods could serve as
a fast and high-precision alternative to first-principles modeling. In the ML framework, predictive computations are modeled as a function from some inputs to the
output of desired values. Because the input data usually come from experiments,
first-principles calculations, or both, applications of ML methods are still limited.
Nevertheless, once input data are collected and the proper ML framework is devised,
ML methods would enable the fast prediction of various kinds of physical properties.
The automation of the process of intuition in materials synthesis (i.e., from idea
to the desired material) is the next step. However, ML methods have just started to
be developed for catalysis due to the difficulty of finding a relationship between the
bulk/surface structure and catalytic performance [7–9]. Although important contributions have recently been made using ML, the synthesis of truly new catalysts using
ML is not currently possible and many aspects of the process remain unexplored.
We have recently described the ML prediction of d-band centers [10]. Even though
there is no theory-based relationship for sufficiently describing heterogeneous catalysis, d-band centers are widely used as a general and versatile descriptor for various
catalytic reactions [11, 12]. However, the accuracy of d-band center model predictions for a given catalytic reaction is relatively low. To develop a more accurate
catalyst design guide and eventually novel catalysts for challenging reactions such
as those that utilize CH 4 , more accurate and specific relationships (descriptors) that
describe catalytic reactions are required.
To meet this challenge, we have followed ML approaches for predicting adsorption
energies. Adsorption is a fundamental step in surface-catalyzed reactions because of
the relation between the heat of adsorption and the activation energy, which follows
the Brønsted–Evans–Polanyi (BEP) relationship [13].
T. Toyao et al.
1 Introduction
Much effort has been devoted to the development of new catalysts for more efficient utilization of CH 4 owing to its economic importance and significant interest in
both industry and academia. The development of such catalysis remains challenging
because the complexity of the surface reactions and the large number of independent
parameters make the prediction of catalyst performance a formidable task. Studies
have recently applied descriptors (indicators or relations) to solid catalysts to find
a relationship between the bulk/surface structure and performance [1]. In principle,
the pace of catalyst discovery could be accelerated by using computational screening
methods [2]. Recent developments in methodology and computer technology, as
well as the establishment of a descriptor-based approach for the analysis of reaction
mechanisms and trends across the periodic table, allow fast screening of new catalytic
materials, which has already resulted in the computational discoveries of several new
materials [3, 4]. However, these computation-based approaches are typically based
on first-principles calculations and thus usually have a high computational cost.
In recent years, machine learning (ML) methods have gained increasing popularity
in the molecular and materials science communities for use in the high-throughput
screening or the prediction of various kinds of physical properties whose mathematical modeling involves complex principles [5, 6]. ML methods could serve as
a fast and high-precision alternative to first-principles modeling. In the ML framework, predictive computations are modeled as a function from some inputs to the
output of desired values. Because the input data usually come from experiments,
first-principles calculations, or both, applications of ML methods are still limited.
Nevertheless, once input data are collected and the proper ML framework is devised,
ML methods would enable the fast prediction of various kinds of physical properties.
The automation of the process of intuition in materials synthesis (i.e., from idea
to the desired material) is the next step. However, ML methods have just started to
be developed for catalysis due to the difficulty of finding a relationship between the
bulk/surface structure and catalytic performance [7–9]. Although important contributions have recently been made using ML, the synthesis of truly new catalysts using
ML is not currently possible and many aspects of the process remain unexplored.
We have recently described the ML prediction of d-band centers [10]. Even though
there is no theory-based relationship for sufficiently describing heterogeneous catalysis, d-band centers are widely used as a general and versatile descriptor for various
catalytic reactions [11, 12]. However, the accuracy of d-band center model predictions for a given catalytic reaction is relatively low. To develop a more accurate
catalyst design guide and eventually novel catalysts for challenging reactions such
as those that utilize CH 4 , more accurate and specific relationships (descriptors) that
describe catalytic reactions are required.
To meet this challenge, we have followed ML approaches for predicting adsorption
energies. Adsorption is a fundamental step in surface-catalyzed reactions because of
the relation between the heat of adsorption and the activation energy, which follows
the Brønsted–Evans–Polanyi (BEP) relationship [13].