Graph Representation of Periodic Systems for Accurate and Explainable Prediction of Material Properties via Machine Learning

Nov 30, 2017 - 2:00 PM -  TC04.09.03
Hynes, Level 2, Room 202
Tian Xie1 , Jeffrey Grossman 1

1 Massachusetts Institute of Technology Cambridge United States
Machine learning (ML) models are being developed with the aim of designing new materials with accuracy close to ab-initio calculations, but speed orders of magnitude faster. Due to the arbitrary size of periodic crystal systems, crystals need to be represented in fixed length to be compatible with most ML algorithms. In existing methods, this problem is resolved by representing crystals with a fixed-length vector converted from atom coordinates under required symmetry invariance. However, one drawback of such an approach is that chemical insights are lacking because of the “black-box” nature of ML models.

In this work, we develop a framework for representing periodic crystal systems by directly building neural networks on top of periodic graphs generated from crystal structures, which provides both material property prediction with DFT accuracy and atomic level chemical insights. The accuracy of our approach is demonstrated by predicting the formation energy, band gap, and fermi energy of inorganic crystals, using calculated data from the open Materials Project database which includes compounds from simple metal oxides to complex minerals. The method shows mean absolute error (MAE) on crystals outside the training set close to the MAE of DFT calculated properties compared to experimental results. More importantly, we demonstrate the ability of this method to provide chemical insights by considering an example of perovskites. We are able to predict the relative stability of different sites in the perovskite structure despite the fact that the model is only trained with total formation energies of the perovskite crystals. The trends given by our model are in good agreement with chemical intuition.