Machine Learning Modeling of Materials with a Group-Subgroup Structure

摘要: Crystal structures connected by continuous phase transitions are related through mathematical relationships between crystallographic groups and their subgroups. In the present study, we introduce group-subgroup machine learning (GS-ML) show that including materials with small unit cells in training set decreases out-of-sample prediction errors for large cells. GS-ML incurs least ML cost to reach 2-3 % target accuracy compared conventional Delta-ML. Since available datasets heterogeneous providing insufficient examples realizing structure, "FriezeRMQ1D" dataset 8393 Q1D organometallic uniformly distributed across seven frieze groups. Furthermore, comparing performances of FCHL 1-hot representations, capture subgroup information only when descriptor encodes structural information. The proposed approach is generic extendable other symmetry abstractions such as spin-, valency-, or charge order.