posted on 2019-07-22, 19:39authored bySeoin Back, Junwoong Yoon, Nianhan Tian, Wen Zhong, Kevin Tran, Zachary W. Ulissi
High-throughput screening
of catalysts can be performed using density
functional theory calculations to predict catalytic properties, often
correlated with adsorbate binding energies. However, more complete
investigations would require an order of 2 more calculations compared
to the current approach, making the computational cost a bottleneck.
Recently developed machine-learning methods have been demonstrated
to predict these properties from hand-crafted features but have struggled
to scale to large composition spaces or complex active sites. Here,
we present an application of a deep-learning convolutional neural
network of atomic surface structures using atomic and Voronoi polyhedra-based
neighbor information. The model effectively learns the most important
surface features to predict binding energies. Our method predicts
CO and H binding energies after training with 12 000 data for
each adsorbate with a mean absolute error of 0.15 eV for a diverse
chemical space. Our method is also capable of creating saliency maps
that determine atomic contributions to binding energies.