Crystal Structure Prediction via Deep Learning

We demonstrate the application of
deep neural networks as a machine-learning
tool for the analysis of a large collection of crystallographic data
contained in the crystal structure repositories. Using input data
in the form of multiperspective atomic fingerprints, which describe
coordination topology around unique crystallographic sites, we show
that the neural-network model can be trained to effectively distinguish
chemical elements based on the topology of their crystallographic
environment. The model also identifies structurally similar atomic
sites in the entire data set of ∼50000 crystal structures,
essentially uncovering trends that reflect the periodic table of elements.
The trained model was used to analyze templates derived from the known
crystal structures in order to predict the likelihood of forming new
compounds that could be generated by placing elements into these structural
templates in a combinatorial fashion. Statistical analysis of predictive
performance of the neural-network model, which was applied to a test
set of structures never seen by the model during training, indicates
its ability to predict known elemental compositions with a high likelihood
of success. In ∼30% of cases, the known compositions were found
among the top 10 most likely candidates proposed by the model. These
results suggest that the approach developed in this work can be used
to effectively guide the synthetic efforts in the discovery of new
materials, especially in the case of systems composed of three or
more chemical elements.