Prediction of Adsorption Energies for Chemical Species on Metal Catalyst Surfaces Using Machine Learning

Computational catalyst
screening has the potential to significantly
accelerate heterogeneous catalyst discovery. Typically, this involves
developing microkinetic reactor models that are based on parameters
obtained from density functional theory and transition-state theory.
To reduce the large computational cost involved in computing various
adsorption and transition-state energies of all possible surface states
on a large number of catalyst models, linear scaling relations for
surface intermediates and transition states have been developed that
only depend on a few, typically one or two descriptors, such as the
carbon atom adsorption energy. As a result, only the descriptor values
have to be computed for various active site models to generate volcano
curves in activity or selectivity. Unfortunately, for more complex
chemistries the predictability of linear scaling relations is unknown.
Also, the selection of descriptors is essentially a trial and error
process. Here, using a database of adsorption energies of the surface
species involved in the decarboxylation and decarbonylation of propionic
acid over eight monometalic transition-metal catalyst surfaces (Ni,
Pt, Pd, Ru, Rh, Re, Cu, Ag), we tested if nonlinear machine learning
(ML) models can outperform the linear scaling relations in prediction
accuracy when predicting the adsorption energy for various species
on a metal surface based on data from the rest of the metal surfaces.
We found linear scaling relations to hold well for predictions across
metals with a mean-absolute error of 0.12 eV, and ML methods being
unable to outperform linear scaling relations when the training dataset
contains a complete set of energies for all of the species on various
metal surfaces. Only when the training dataset is incomplete, namely,
contains a random subset of species’ energies for each metal,
a currently unlikely scenario for catalyst screening, do kernel-based
ML models significantly outperform linear scaling relations. We also
found that simple coordinate-free species descriptors, such as bond
counts, achieve as good results as sophisticated coordinate-based
descriptors. Finally, we propose an approach for automatic discovery
of appropriate metal descriptors using principal component analysis.