STEM2Crystal-Bench

STEM2Crystal-Bench is the benchmark for the paper "From Noisy STEM to Crystal Structure: Evidence-Structure CoDiffusion under Composition Constraints" (Chen & You, KDD 2026, Oral), which introduces STEM2Crystal CoDiffusion (SCCD). It evaluates methods that reconstruct a crystal structure from a noisy STEM image when the composition is known. The release has a large synthetic set with controlled noise and a small set of real STEM images, with ground-truth CIFs and a fixed test split.

Code and a live leaderboard: https://github.com/PEESEgroup/STEM2Crystal-Bench

Contents

The release has 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, open-licensed atomic-resolution STEM images of 2D monolayer materials, paired with single-layer (monolayer) ground-truth CIFs	Out-of-distribution real-microscopy probe (paper §5.4, Table 2)

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 are all genuine 2D monolayer materials (4 transition-metal dichalcogenides spanning Mo/W × S/Se/Te, plus graphene), assembled as an out-of-distribution real-microscopy probe. Each image is redistributed here only because the original source is open-licensed (CC BY 4.0 or CC BY 3.0); the ground truth is the single-layer (monolayer) structure, so a 2D-slab model can match it without prototype templating. We explicitly acknowledge the original authors below.

#	sid	Material	Phase	Image source	Source paper	DOI	License
1	real01_MoTe2_monolayer	monolayer 2H-MoTe₂	hexagonal P-6m2 (monolayer)	Zenodo 5520169	Guo et al., npj Comput Mater (2021)	10.1038/s41524-021-00642-1	CC BY 4.0
2	real02_WSe2_monolayer	monolayer 2H-WSe₂	hexagonal P-6m2 (monolayer)	Zenodo 7696721	Tan et al., npj Comput Mater (2023)	10.1038/s41524-023-01042-3	CC BY 4.0
3	real03_WS2_monolayer	monolayer 2H-WS₂ (+S vacancies)	hexagonal P-6m2 (monolayer)	figshare 25310755	Wang et al., Nat Commun (2024)	10.1038/s41467-024-53880-4	CC BY 4.0
4	real04_MoS2_monolayer	monolayer 1H-MoS₂	hexagonal P-6m2	Wikimedia Commons MoS2_vacancies.jpg	Hong et al., Nat Commun 6 (2015)	10.1038/ncomms7293	CC BY 4.0
5	real05_graphene	monolayer graphene	honeycomb P6/mmm	Wikimedia Commons Graphene-TEM.jpg	Gass et al., Nat Nanotechnol 3 (2008)	10.1038/nnano.2008.280	CC BY 3.0

Ground-truth CIFs are the single-layer (monolayer) structures (TMD layers derived from their COD bulk entries; MoS₂ and graphene use the canonical monolayer cells). Source paper PDFs are not redistributed; please consult the DOIs above.

Citation

If you use STEM2Crystal-Bench, please cite our paper (KDD 2026, Oral) and the original source papers for the real_stem_eval5 images listed above.

@inproceedings{chen2026stem2crystal,
  title     = {From Noisy STEM to Crystal Structure: Evidence-Structure CoDiffusion under Composition Constraints},
  author    = {Chen, Guangyao and You, Fengqi},
  booktitle = {Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD)},
  year      = {2026},
  note      = {Oral presentation}
}

Code + live leaderboard: https://github.com/PEESEgroup/STEM2Crystal-Bench

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/: monolayer structures): upstream licenses (C2DB is CC BY 4.0, MC2D is CC BY 4.0 via Materials Cloud, COD is public-domain)
• Real STEM images (real_stem_eval5/images/): CC BY 4.0 (four TMD monolayers) and CC BY 3.0 (graphene), redistributed under the terms of the original open-license sources (Zenodo / figshare / Wikimedia Commons), with full attribution to the source authors in the table above