講演情報
[10a-N302-2]From Chemical Formula to Crystal Structure: Space-Group Prediction Guides Template Retrieval
〇YenJu Wu1, Yibin Xu1 (1.NIMS)
キーワード：
Structure prediction、Space group、Structure template retrieval
Generative machine learning can propose novel chemical compositions, but these candidates often lack crystal structures required for simulation and property evaluation. Here, we develop a formula-to-structure pipeline that uses predicted space-group information to guide template-based structure generation. A chemical formula is converted into periodic descriptors, candidate space groups are predicted using XGBoost, and existing crystal structures are retrieved as templates from the predicted space-group families. The retrieved templates are then converted to the target composition by stoichiometry-aware element substitution and relaxed using MatterSim.We compare space-group-guided retrieval with unconstrained descriptor-based retrieval using leave-one-out benchmarks on AtomWork-Adv. and Materials Project datasets. Restricting templates to the top-1 predicted space group improves the space-group match rate of relaxed structures from 44% to 72% on AtomWork-Adv. and from 32% to 57% on Materials Project. These results show that predicted space-group information provides an effective and interpretable constraint for selecting better structure templates from chemical formulas, offering a practical bridge between generative composition models and physically meaningful crystal structures.