In materials discovery applications often we know the composition of trial materials but have little knowledge about the structure. Many current SOTA results within the field of machine learning for materials discovery are reliant on knowledge of the structure of the material. This means that such models can only be applied to systems that have undergone structural characterisation. As structural characterisation is a time-consuming process whether done experimentally or via the use of ab-initio methods the use of structures as our model inputs is a prohibitive bottleneck to many materials screening applications we would like to pursue. One approach for avoiding the structure bottleneck is to develop models that learn from the stoichiometry alone. In this work, we show that via a novel recasting of how we view the stoichiometry of a material we can leverage a message-passing neural network to learn materials properties whilst remaining agnostic to the structure. The proposed model exhibits increased sample efficiency compared to more widely used descriptor-based approaches. This work draws inspiration from recent progress in using graph-based methods for the study of small molecules and crystalline materials.
Goodall, R.E.A., Lee, A.A. Predicting materials properties without crystal structure: deep representation learning from stoichiometry. Nat Commun 11, 6280 (2020).
Stoichiometry, Representation Learning, Materials Science,
Curator: qinyang