Material Property Prediction Using Graphs Based on Generically Complete Isometry Invariants

• DOI: 10.1007/s40192-024-00351-9
• Publication Date: 2024-04-16T15:28:39Z
• Authors: Jonathan Balasingham; Viktor Zamaraev; Vitaliy Kurlin
• Keywords: Pointwise; Crystal (programming language)

Abstract The structure–property hypothesis says that the properties of all materials are determined by an underlying crystal structure. main obstacle was ambiguity conventional representations based on incomplete or discontinuous descriptors allow false negatives positives. This resolved ultra-fast pointwise distance distribution, which distinguished periodic structures in world’s largest collection real (Cambridge structural database). State-of-the-art results property prediction were previously achieved graph neural networks various crystals, including Crystal Graph with vertices at atoms a unit cell. work adapts distribution for simpler whose vertex set is not larger than asymmetric new Distribution reduces mean absolute error 0.6–12% while having 44–88% number when compared to applied Materials Project and Jarvis-DFT datasets using CGCNN ALIGNN. Methods hyper-parameters selection backed theoretical then experimentally justified.