STEM2Crystal-Bench

STEM2Crystal-Bench is the benchmark accompanying the paper "From Noisy STEM to Crystal Structure: Evidence-Structure CoDiffusion under Composition Constraints" (Chen & You, 2026), which introduces STEM2Crystal CoDiffusion (SCCD). STEM2Crystal-Bench is a controlled large-scale benchmark with explicit composition constraints, controlled noise regimes, and projection-derived supervision signals, designed for reproducible evaluation of microscopy-guided crystal structure prediction from noisy STEM observations.

Contents

This release exposes two test configs:

Config	#samples	Source	Role
synthetic	1021	Physics-inspired forward pipeline (paper Appendix A.3) applied to curated 2D slab prototypes from C2DB + MC2D	Primary evaluation used in paper Table 1 under three noise regimes (low / mid / high)
real_stem_eval5	5	Real experimental atomic-resolution HAADF / iDPC-STEM images from open-license deposits, paired with ground-truth CIFs from the Crystallography Open Database (COD)	Out-of-distribution real-microscopy probe (added in rebuttal)

Folder structure

STEM2Crystal-Bench/
├── README.md
├── LICENSE                        # CC BY 4.0
├── synthetic/
│   ├── cifs/                      # 1021 ground-truth CIFs (C2DB + MC2D)
│   ├── images/
│   │   ├── low/                   # 3063 images (1021×3 views), low shot noise
│   │   ├── mid/                   # 3063 images, medium noise
│   │   └── high/                  # 3063 images, high noise
│   ├── masks/                     # atom-position target masks
│   ├── split.json                 # test split IDs (same as paper §5)
│   └── metadata.jsonl             # per-sample lattice/formula/SG
└── real_stem_eval5/
    ├── cifs/                      # 5 COD ground-truth CIFs
    ├── images/                    # 5 real experimental STEM PNGs
    └── metadata.jsonl             # per-sample with source attribution

All images in synthetic/ are resized to 256×256 (the input size used by SCCD and baselines). The full-resolution abTEM output (1286×866 per image, ≈88 GB total) is not redistributed here; the simulation recipe is released with the paper's code.

Loading

from datasets import load_dataset

synth = load_dataset("gary23ai/STEM2Crystal-Bench", "synthetic", split="test")
real  = load_dataset("gary23ai/STEM2Crystal-Bench", "real_stem_eval5", split="test")

Each sample exposes mp_id, formula, lattice parameters (a, b, c, alpha, beta, gamma), space_group, and per-noise image paths.

real_stem_eval5/ config — data sources

The 5 real experimental samples were assembled specifically as an out-of-distribution real-microscopy probe for the rebuttal of this paper. Every image is redistributed here only because the original deposit is open-licensed (CC BY 4.0). We explicitly acknowledge the original authors and papers below.

#	sid	Material	Phase	Source deposit	Source paper	DOI	COD entry
1	real01_ZrO2_monoclinic	ZrO₂	monoclinic P2₁/c (baddeleyite)	Zenodo 5520169	Guo et al., npj Comput Mater 7 (2021)	10.1038/s41524-021-00527-3	COD 1010912
2	real02_2H_MoTe2	2H-(Mo,W)Te₂	hexagonal P6₃/mmc	Zenodo 5520169	Guo et al., npj Comput Mater 7 (2021)	10.1038/s41524-021-00527-3	COD 2310465
3	real03_Al_FCC	Al	FCC Fm-3m + partial dislocations	Zenodo 5520169	Guo et al., npj Comput Mater 7 (2021)	10.1038/s41524-021-00527-3	COD 1502689
4	real04_2H_WSe2_monolayer	monolayer 2H-WSe₂	hexagonal P6₃/mmc	Zenodo 7696721	Tan et al., npj Comput Mater 9 (2023)	10.1038/s41524-023-01042-3	COD 9012193
5	real05_2H_WS2_monolayer	monolayer 2H-WS₂ + S-vacancies	hexagonal P6₃/mmc	figshare 25310755	Wang et al., Nat Commun 15 (2024)	10.1038/s41467-024-48069-8	COD 9012191

All ground-truth CIFs are from the Crystallography Open Database (COD), public-domain. Source paper PDFs are not redistributed here; please consult the DOIs above for the original publications.

Citation

If you use STEM2Crystal-Bench or the Real-STEM-Eval5 subset, please cite our paper and the original source papers for real_stem_eval5 listed above.

@inproceedings{sccd2026,
  title={From Noisy STEM to Crystal Structure: Evidence-Structure
         CoDiffusion under Composition Constraints},
  author={Chen, Guangyao and You, Fengqi},
  year={2026},
  note={Submission under review}
}

License

• Benchmark assembly, metadata, splits, synthetic STEM renderings: CC BY 4.0
• Ground-truth CIFs (synthetic/cifs/: 764 from C2DB + 257 from MC2D; real_stem_eval5/cifs/ from COD): upstream licenses (C2DB is CC BY 4.0, MC2D is CC BY 4.0 via Materials Cloud, COD is public-domain)
• Real experimental STEM images (real_stem_eval5/images/): CC BY 4.0, redistributed under the terms of the original Zenodo / figshare deposits, with full attribution to the source authors in the table above