Machine Learning Prediction of CO Adsorption Energies and Properties of Layered Alloys Using an Improved Feature Selection Algorithm

Layered alloys are widely studied
as designable catalysts. As a
surface probe molecule, CO adsorption energy is not only employed
to characterize surface properties but also used as the catalytic
activity descriptor in various reactions. With the aid of high-throughput
computing technology, we calculated CO adsorption energies on 3729layered
alloy surfaces. To obtain CO adsorption energies, the d-band center
and d-band skewness, and the stability of all the remaining layered
alloys (8415) of 23 transition metals, we collected 91 features that
do not require time-consuming quantum chemistry calculations (non-QC
features) and 40 features from quantum chemistry calculations (QC
features). To reduce the feature dimension and overcome overfitting
problems, we proposed a modified sequential feature selection (SFS)
wrapper method to identify (sub)optimal subsets. Two supervised light
gradient boosting machine regression (LGBMR) machine learning (ML)
regression models were established using the identified subsets. It
is demonstrated that the size of the feature subset converges rapidly,
and the performance of the model with size nine is already quite satisfactory.
The ML model of the non-QC features outperforms that of QC features.
Using the ML models established with non-QC features, we predicted
the CO adsorption energies and the electronic structure–properties
(d-band center, d-band skewness) and stability of 8415 layered alloys.
Based on the four conditions (CO adsorption energy, stability, price,
and surface segregation), potential alloy catalysts for CO2 to methanol were screened out of 12144 layered alloys.