Encoding chemical similarity using pairwise distances.

13^th May 2022, 14:00
James Cumby
University of Edinburgh

Abstract

In order to design new materials that fulfil a desired function, we require tools to rapidly predict physical properties from the underlying atomic structure. Statistical machine learning is accelerating advances in this area, but a key limitation remains; accurately predicting properties requires an effective measure of similarity between crystal structures. Here, we show that a representation based on grouped pairwise inter-atomic distances (GRID) combined with earth mover's distance (EMD) gives an accurate reflection of chemical space, and can be used to predict multiple physical properties using a simple regression model.