SymmRelax processed binary-crystal graph dataset, trained checkpoints and source data

This record contains the processed dataset, trained model checkpoints and source data associated with the manuscript “Symmetry preserving pre-relaxation for high-throughput and reliable crystal screening”. The processed dataset is provided as processed_sqlite.tar.xz and contains the SQLite graph dataset, metadata, preprocessing summary and train/validation/test split files used for SymmRelax training, inference and evaluation. The archive includes data.sqlite, meta.csv, skip_summary.csv, split_train.npy, split_val.npy and split_test.npy. The record also includes the trained SymmRelax geometry checkpoint, the trained CGCNN/CGCNN-EAF checkpoint used for downstream or optional energy-prediction workflows, and SourceData.zip containing the source data underlying the figures, tables, Extended Data and selected Supplementary Tables. The original crystal entries were collected from publicly available materials databases, including Alexandria, the Materials Project, OQMD and AFLOW, as described in the manuscript. The source code, preprocessing workflow, training scripts and inference scripts are available at https://github.com/zrfcms-buaa/SymmRelax.