Labeled Synthetic Crystal Material Dataset

Overview The LSCM (Labeled Synthetic Crystal Materials) dataset contains crystal structures generated by CrystaLLM code and labeled with seven pre-trained foundation models. There are a total of 7,889 distinct synthetic crystalline structures in the database. The uniqueness of each structure, i.e., whether it is a known structure or exists in the training dataset for CrystaLLM, has not been checked against. The database is an ASE database in SQLite3 format. For details on how to read the database using ASE, refer to the ASE documentation. Each structure produced by CrystaLLM is validated for spacegroup symmetry compliance and relaxed using symmetry-constrained optimization for a given pre-trained foundation model. The database labels include energy, forces, and stress provided by the foundation model. Reading the Database with ASE from ase.db import connect db = connect('LSCM_Relaxed_10July2025.db') Foundation Models Calculator Engine (Model Checkpoint) Description Orb v3-conservative-inf-omat ORB v3 conservative model with infinite cutoff sevenn 7net-mf-ompa evenNet MF OMPA variant mace mace-mpa-0 MACE MPA-0 model mattersim v1-5m MatterSim v1 (5M parameters) deepmd dpa-3.1-3m-ft DPA-3.1, 3M parameters, finetuned eqnorm mptrj Equivariant Normalizing Flow model (MPTraj variant) grace 2l-oam GRACE 2-layer OAM To view the respective calculator and model details, check matbench discovery project page or references. DisclaimerThis dataset was generated using the author’s personal computing resources and during independent research activities conducted completly outside or prior to the scope of any employment or institutional affiliation. The dataset and any associated analyses here have not been peer-reviewed. Availability of this material does not imply accuracy, completeness, validation, or fitness for any particular purpose.