To
realize a renewable and sustainable energy cycle, there has
been a lot of effort put into discovering catalysts with desired properties
from a large chemical space. To achieve this goal, several screening
strategies have been proposed, most of which require validation of
thermodynamic stability and synthesizability of candidate materials
via computationally intensive quantum chemistry or solid-state physics
calculations. This problem can be overcome by reducing the number
of calculations through machine learning methods, which predict target
properties using unrelaxed crystal structures as inputs. However,
numerical input representations of most of the previous models are
based on either too specific (e.g., atomic coordinates) or too ambiguous
(e.g., stoichiometry) information, practically inapplicable to energy
prediction of unrelaxed initial structures. In this work, we develop
a direction-based crystal graph convolutional neural network (D-CGCNN)
with the highest accuracy toward formation energy predictions of the
relaxed structures using the initial structures as inputs. By comparing
with other approaches, we revealed correlations between crystal graph
similarities and model performances, elucidating the origin of the
improved accuracy of our model. We applied this model to the ongoing
high-throughput virtual screening project, where the model discovered
1,725 stable materials from 15,318 unrelaxed structures by performing
3,966 structure optimizations (∼25%).