STEM2Crystal-Bench — microscopy-guided crystal structure reconstruction

The benchmark

The long-term aim is a direct bridge from microscopy to materials: generate a structure straight from a STEM image, then analyze it. This benchmark targets the first and hardest link, recovering a crystal structure from a single noisy image. It pairs STEM images with crystal targets and one evaluation suite, in two parts:

Synthetic

1021 structures · three controlled noise regimes (low / mid / high). Physics‑inspired forward model: probe blur → Poisson counting → scan jitter → background → readout.

Real‑5

5 genuine 2D‑monolayer real STEM images (MoTe₂, WSe₂, WS₂, MoS₂, graphene) with single‑layer ground truth. Zero‑shot, no templating.

Data lives on the Hugging Face Hub (not in the code repo): python scripts/download_benchmark.py --config all

Reading a structure out of a microscopy image is still an open problem. The benchmark is meant to grow: contributions of data (other material classes, more real images) and of new methods are welcome.

Leaderboard

↑ higher is better · ↓ lower is better · best per column in bold. Updated — · read live from GitHub.

Qualitative comparison

Each method's rank-1 reconstruction next to the ground truth, labelled with the RMS to GT.

reconstruction comparison across methods

Regenerate from predictions with scripts/plot_comparison.py (needs pip install -e ".[viz]").

Methods

Each method implements the same Method interface and is scored the same way. Click one for details, or add your own.

Evaluate and contribute

pip install -e .
python scripts/download_benchmark.py --config all
python scripts/evaluate.py --method <name> --benchmark synthetic --noise low   # all paper metrics

Contributions are welcome: a method (through the Method interface, with a row added to leaderboard.json), a new metric (a pure function in eval/metrics.py), or new data (other material classes or more real STEM images) on the dataset page.

Citation

If you use STEM2Crystal-Bench or SCCD, please cite: