Abstract

Coarse-graining to molecular graphs is critical to ML-driven molecular discovery. We show how Wyckoff representations – coordinate-free sets of symmetry-related positions in crystals – can empower analogous improvements in inorganic material discovery. Unlike atomic positions, Wyckoff representations are computably enumerable enabling screening campaigns across novel materials space. Critically, compared to composition-based approaches we avoid the majority of the cost of structure searching when validating predictions. We propose “Wren” – a ML model operating on Wyckoff representations. We demonstrate Wren’s potential by identifying 1,558 novel materials that lie below the convex hull of previously calculated materials from just 5,675 ab-initio calculations. arXiv pre-print: https://arxiv.org/abs/2106.11132