STEMGym: Benchmark Data for Dose-Efficient Autonomous STEM Microscopy

This dataset accompanies the STEMGym benchmark — a Gymnasium-based environment for evaluating autonomous dose-efficient scanning transmission electron microscopy (STEM) agents.

Dataset Description

STEMGym provides simulated STEM specimens as HDF5 world files. Each world contains:

• Overview image: Low-magnification survey of the full specimen
• Tile grid: High-resolution STEM images (128×128 px tiles, 4px overlap, stride=124) arranged in an 8×8 grid
• Ground truth annotations: Atom positions, defect types, and phase maps for scoring
• Metadata: Pixel size, accelerating voltage, detector geometry, material parameters

Materials

Material	Description	Defect Types
SrTiO₃	Perovskite oxide	O vacancies, Sr vacancies, Ti antisites
BaTiO₃	Ferroelectric perovskite	O vacancies, Ba vacancies, domain boundaries
SiGe	Semiconductor alloy	Ge substitutions, strain fields

Difficulty Levels

Each material is provided at three difficulty levels (easy, medium, hard) controlling defect density, noise level, and spatial distribution.

Files

Simulated Worlds (worlds/)

File	Material	Difficulty	Size
test_world.h5	Synthetic (Gaussian blobs)	—	~5 MB
srtio3_easy.h5	SrTiO₃	Easy	~50 MB
srtio3_medium.h5	SrTiO₃	Medium	~50 MB
srtio3_hard.h5	SrTiO₃	Hard	~50 MB
batio3_easy.h5	BaTiO₃	Easy	~50 MB
batio3_medium.h5	BaTiO₃	Medium	~50 MB
batio3_hard.h5	BaTiO₃	Hard	~50 MB
sige_easy.h5	SiGe	Easy	~50 MB
sige_medium.h5	SiGe	Medium	~50 MB
sige_hard.h5	SiGe	Hard	~50 MB

Model Checkpoints (checkpoints/)

File	Model	Description	Size
atom_finder.pt	AtomFinderUNet	Atomic column detection ensemble (3 members)	~88 MB
defect_classifier.pt	DefectClassifierCNN	Defect type classification	~1.4 MB
phase_identifier.pt	PhaseIdentifierResNet	Material phase identification	~7.4 MB

HDF5 World Format

Each world file follows this layout:

/metadata/
    pixel_size_nm          # Physical pixel size
    tile_size_px           # Tile dimensions (128)
    tile_overlap_px        # Overlap between tiles (4)
    grid_rows, grid_cols   # Grid dimensions
    accelerating_voltage   # Beam energy (keV)
    convergence_angle      # Probe convergence (mrad)
    detector_inner/outer   # HAADF detector angles (mrad)
/overview                  # Low-magnification survey image
/tiles/{row}_{col}         # High-resolution tile images
/ground_truth/
    atom_positions         # (N, 3) array: x, y, atomic_number
    defect_types           # (M, 4) array: x, y, type_id, severity
    phase_map              # 2D array: phase labels per pixel
/valid_region              # Boolean mask of valid scan area

Usage

# Install STEMGym
pip install -e .

# Download this data
python scripts/download_data.py

# Run a benchmark
stemgym run --agent raster_equipped --task defect_census --world srtio3_medium --seeds 3

License

This dataset is released under the CC-BY-4.0 license.

Citation

@inproceedings{stemgym2026,
  title={STEMGym: A Benchmark for Dose-Efficient Autonomous STEM Microscopy},
  author={Polat, John},
  year={2026}
}