Binary Compound Crystal Structures for Energy Relaxation Modeling

Paired serialized CIF records for perturbed binary compound structures and their corresponding DFT-relaxed targets. The dataset is organized into separate train and test source-target text files for the CrystalRelax task. It was created by Xu and released on the figshare platform in April 2026.

Use Cases

• Train language models to generate relaxed crystal structures from perturbed inputs using the serialized CIF text pairs.
• Evaluate model performance on the CrystalRelax task using the provided src-test and tgt-test files for CIF validity and per-atom potential energy comparison.
• Benchmark reinforcement learning agents for accelerating crystal structure relaxation by using generated structures as lower-energy starting points.
• Fine-tune models on CIF-format generation and energy-aware structural relaxation for binary compounds.

Strengths

• Provides paired perturbed and DFT-relaxed structures, enabling supervised learning for a specific physical task.
• Dataset is structured into distinct train and test splits (src-train.txt, tgt-train.txt, src-test.txt, tgt-test.txt) for direct model evaluation.
• Released under a permissive CC-BY-4.0 license for open research use.

Limitations

• The dataset is small at 8.7 MB, indicating a limited number of structure pairs, which may constrain model training.
• Limited to binary compounds, reducing generalizability to more complex multi-element crystal systems.
• Specific row count and the diversity of chemical species covered are unknown.

Provenance

Source

figshare, uploaded by author Xu.

Collection Method

Structures were generated and relaxed using Density Functional Theory (DFT) methods, as indicated by 'DFT-relaxed target structures'.

Freshness

Last updated in April 2026, indicating recent creation.