STEM2Crystal-Bench: Noisy STEM Images for Crystal Structure Reconstruction

STEM2Crystal-Bench is a benchmark for evaluating methods that reconstruct crystal structures from noisy scanning transmission electron microscopy images when the composition is known. The dataset includes a large synthetic set with controlled noise and a smaller set of real STEM images, both paired with ground-truth crystal information files. It was created by author gary23ai and released in 2026.

Use Cases

• Benchmarking crystal structure reconstruction models based on noisy STEM images.
• Training denoising or generative models based on the synthetic dataset with controlled noise.
• Evaluating the generalization of models from synthetic to real data based on the included real STEM images.
• Developing multimodal models that combine image and compositional constraints for materials discovery.

Strengths

• Includes both a large synthetic dataset and a smaller set of real STEM images, providing a controlled testbed and a real-world evaluation component.
• Provides ground-truth crystal information files for validation.
• Serves as the official benchmark for a KDD 2026 Oral paper, indicating a research-grade standard.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count and overall dataset size are unknown, which may limit suitability assessment.
• The description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

huggingface

Collection Method

Likely contains synthetic data generated with controlled noise and a collection of real STEM images.

Time Range

null

Freshness

Last updated 2026-06-11 16:41:22; freshness should be verified.

Geography

null