Machine learning models recently demonstrated prominent results in predicting materials properties and accelerating materials discovery [1]. However, they still fail in predicting complex material properties (particularly bandgaps) or maintaining accuracy in the low-data regime.

Graph Convolution Neural Networks (GCNNs) became the de facto standard for encoding crystal structures of arbitrary size. They showed high accuracy in predicting additive properties of the materials, e.g. total energy, which can be represented as a sum of energies of single atoms. However, we find that they fail to predict atomic interactions that result in delocalized properties, such as molecular orbitals and the resulting electronic bandstructure, even in the simplest 2-level system case.

Here, we developed a material representation that retains the locality of atomic contributions to the final molecular orbitals and band structure. Our GCNN trained on band structure data from the Materials Project database [2] with elements from s, p, d, and f blocks is capable of making accurate predictions of projected densities of states even when training on a dataset with less than 5000 materials.

[1] Xie, T. and Grossman, J. C., “Crystal Graph Convolutional Neural Networks for an Accurate and Interpretable Prediction of Material Properties”, Physical Review Letters, vol. 120, no. 14, 2018. doi:10.1103/PhysRevLett.120.145301.

[2] Anubhav Jain, Shyue Ping Ong, Geoffroy Hautier, Wei Chen, William Davidson Richards, Stephen Dacek, Shreyas Cholia, Dan Gunter, David Skinner, Gerbrand Ceder, and Kristin A. Persson, "Commentary: The Materials Project: A materials genome approach to accelerating materials innovation", APL Materials 1, 011002 (2013) https://doi.org/10.1063/1.4812323