The Catalyst Review May 2020 - 8

SPECIAL FEATURE I
Case Study 2: "Rapid Weeding" Surrogate Models for Studying Catalytic Surface Reaction Networks

Cost is a key consideration in the use of computational modelling in catalysis. The number of permutations of reaction pathways,
catalyst surfaces and intermediates can run into trillions, even for simple chemical systems. Therefore, a major goal for researchers
has been to rationalize methods and reduce the computational time and costs. Prof Zachary Ulissi together with researchers at the
SUNCAT Center for Interface Science and Catalysis, a joint SLAC/Stanford institute, developed new machine learning methods to do
just that (Ulissi et al. 2017).
The ground-breaking method uses a novel machine learning system which analyzes the reagents in a specific reaction and, from a
table of known properties, selects the catalysts with the greatest potential performance. The benefit is that only a small number of
initial DFT calculations are required to train the model. Thereafter, it can begin to predict the properties of new surfaces and cuts
out the need for expensive full DFT calculations where they are not needed. The method, which the researchers refer to as "rapid
weeding," cuts through an enormous forest of potential pathways to provide a small number of important "pruned" pathways for
further examination.
Using syngas conversion to ethanol over rhodium (III) species as the system, the team devised a surrogate model which involves
training an algorithm based on a set of existing DFT calculations of surface intermediate formation energies. Through an iterative
process of feedback and prediction, the method allows selection of only those reaction pathways likely to represent the ratelimiting steps for further analysis. These pathways were then analyzed using more detailed and expensive Bayesian error estimation
functional with van der Waals correlation (BEEF-vdW) methods. This machine learning approach saved time and effort by filtering out
less likely reaction pathways. The authors demonstrated reductions of 60% for intermediate and 95% for transition-state calculations,
respectively.
In later researched published at Carnegie Mellon, the method was applied to screen catalysts for the electrochemical CO2 reduction
reaction (CO2RR) and the H2 evolution reaction (HER). A constrained database of surface reactivity of metals is allied with a pipeline
optimisation tool to train the system using enough representative examples. Combinations of 31 elements were included in the
screening stage and the machine learning system identified 131 candidate surfaces across 54 alloys for CO2RR and 258 candidate
surfaces across 102 alloys for HER. A heuristic method was applied to filter these candidates providing a shortlist of 10 candidates
for catalysing CO2RR and 14 for HER (Tran and Ulissi 2018). Effectively this screening takes the time for making catalyst improvements
down from 1-2 years to a matter of weeks or months.
As with any learning system, training must be complemented by feedback. In this example the performance descriptors are the
adsorption energies of *CO for the CO2 reduction reaction and *H for the H2 evolution reaction. Further refinement of the system is
provided via validation of the highest potential candidate catalysts by solution of DFT calculations and feeding the results back into
the model.
The researchers stress that this development is at an early stage, but it promises enormous benefits by allowing cross-disciplinary
access for a larger number of researchers to a simplified and demystified methodology to what has traditionally been the reserve of a
narrow band of skilled material properties experts.
Looking Ahead

Many challenges still lie ahead for AI and machine learning in catalysis:
*

Modelling reactions through a full catalytic cycle; here, the challenge is obtaining enough relevant training data for bondbreaking reactions.

Another challenge is the interpolation of bond-breaking events (compared to non-bond-breaking events) due to the more
significant change in chemical properties of the system.

A third challenge is the difficulty of generating accurate training data for condensed‐phase systems, which contain at least four
different elements.

One thing is almost certain, and that is that these challenges will be met and overcome, now that the winter has ended, and the
Spring of AI is underway.
References

Li Z, Wang S and Xin H. (2018). "Toward Artificial Intelligence in Catalysis." Nature Catalysis, 1: 641-642.
MathWorks. (2016). "Introducing Machine Learning." https://uk.mathworks.com/content/dam/mathworks/tag-team/Objects/
i/88174_92991v00_machine_learning_section1_ebook.pdf
8

The Catalyst Review

May 2020

https://uk.mathworks.com/content/dam/mathworks/tag-team/Objects/i/88174_92991v00_machine_learning_section1_ebook.pdf https://uk.mathworks.com/content/dam/mathworks/tag-team/Objects/i/88174_92991v00_machine_learning_section1_ebook.pdf

The Catalyst Review May 2020

Table of Contents for the Digital Edition of The Catalyst Review May 2020