johnpolat commited on
Commit
eaa3a9b
·
verified ·
1 Parent(s): 287e6ba

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +66 -41
README.md CHANGED
@@ -9,6 +9,7 @@ tags:
9
  - gymnasium
10
  - benchmark
11
  - dose-efficiency
 
12
  pretty_name: STEMGym
13
  size_categories:
14
  - 1K<n<10K
@@ -16,28 +17,36 @@ size_categories:
16
 
17
  # STEMGym: Benchmark Data for Dose-Efficient Autonomous STEM Microscopy
18
 
19
- This dataset accompanies the [STEMGym](https://github.com/KurbanIntelligenceLab/STEMGym) benchmark — a Gymnasium-based environment for evaluating autonomous dose-efficient scanning transmission electron microscopy (STEM) agents.
20
 
21
  ## Dataset Description
22
 
23
  STEMGym provides simulated STEM specimens as HDF5 world files. Each world contains:
24
 
25
  - **Overview image**: Low-magnification survey of the full specimen
26
- - **Tile grid**: High-resolution STEM images (128×128 px tiles, 4px overlap, stride=124) arranged in an 8×8 grid
27
  - **Ground truth annotations**: Atom positions, defect types, and phase maps for scoring
28
  - **Metadata**: Pixel size, accelerating voltage, detector geometry, material parameters
29
 
30
  ### Materials
31
 
32
- | Material | Description | Defect Types |
33
- |----------|-------------|-------------|
34
- | SrTiO₃ | Perovskite oxide | O vacancies, Sr vacancies, Ti antisites |
35
- | BaTiO₃ | Ferroelectric perovskite | O vacancies, Ba vacancies, domain boundaries |
36
- | SiGe | Semiconductor alloy | Ge substitutions, strain fields |
 
 
37
 
38
  ### Difficulty Levels
39
 
40
- Each material is provided at three difficulty levels (easy, medium, hard) controlling defect density, noise level, and spatial distribution.
 
 
 
 
 
 
41
 
42
  ## Files
43
 
@@ -45,24 +54,45 @@ Each material is provided at three difficulty levels (easy, medium, hard) contro
45
 
46
  | File | Material | Difficulty | Size |
47
  |------|----------|-----------|------|
48
- | `test_world.h5` | Synthetic (Gaussian blobs) | — | ~5 MB |
49
- | `srtio3_easy.h5` | SrTiO₃ | Easy | ~50 MB |
50
- | `srtio3_medium.h5` | SrTiO₃ | Medium | ~50 MB |
51
- | `srtio3_hard.h5` | SrTiO₃ | Hard | ~50 MB |
52
- | `batio3_easy.h5` | BaTiO₃ | Easy | ~50 MB |
53
- | `batio3_medium.h5` | BaTiO₃ | Medium | ~50 MB |
54
- | `batio3_hard.h5` | BaTiO₃ | Hard | ~50 MB |
55
- | `sige_easy.h5` | SiGe | Easy | ~50 MB |
56
- | `sige_medium.h5` | SiGe | Medium | ~50 MB |
57
- | `sige_hard.h5` | SiGe | Hard | ~50 MB |
 
 
 
 
 
 
 
58
 
59
  ### Model Checkpoints (`checkpoints/`)
60
 
61
  | File | Model | Description | Size |
62
  |------|-------|-------------|------|
63
- | `atom_finder.pt` | AtomFinderUNet | Atomic column detection ensemble (3 members) | ~88 MB |
64
- | `defect_classifier.pt` | DefectClassifierCNN | Defect type classification | ~1.4 MB |
65
- | `phase_identifier.pt` | PhaseIdentifierResNet | Material phase identification | ~7.4 MB |
 
 
 
 
 
 
 
 
 
 
 
 
 
 
66
 
67
  ## HDF5 World Format
68
 
@@ -72,18 +102,19 @@ Each world file follows this layout:
72
  /metadata/
73
  pixel_size_nm # Physical pixel size
74
  tile_size_px # Tile dimensions (128)
75
- tile_overlap_px # Overlap between tiles (4)
76
- grid_rows, grid_cols # Grid dimensions
77
- accelerating_voltage # Beam energy (keV)
78
- convergence_angle # Probe convergence (mrad)
79
- detector_inner/outer # HAADF detector angles (mrad)
80
- /overview # Low-magnification survey image
81
- /tiles/{row}_{col} # High-resolution tile images
82
  /ground_truth/
83
- atom_positions # (N, 3) array: x, y, atomic_number
84
- defect_types # (M, 4) array: x, y, type_id, severity
85
- phase_map # 2D array: phase labels per pixel
86
- /valid_region # Boolean mask of valid scan area
 
 
87
  ```
88
 
89
  ## Usage
@@ -93,10 +124,10 @@ Each world file follows this layout:
93
  pip install -e .
94
 
95
  # Download this data
96
- python scripts/download_data.py
97
 
98
  # Run a benchmark
99
- stemgym run --agent raster_equipped --task defect_census --world srtio3_medium --seeds 3
100
  ```
101
 
102
  ## License
@@ -105,10 +136,4 @@ This dataset is released under the [CC-BY-4.0](https://creativecommons.org/licen
105
 
106
  ## Citation
107
 
108
- ```bibtex
109
- @inproceedings{stemgym2026,
110
- title={STEMGym: A Benchmark for Dose-Efficient Autonomous STEM Microscopy},
111
- author={Polat, John},
112
- year={2026}
113
- }
114
- ```
 
9
  - gymnasium
10
  - benchmark
11
  - dose-efficiency
12
+ - autonomous-microscopy
13
  pretty_name: STEMGym
14
  size_categories:
15
  - 1K<n<10K
 
17
 
18
  # STEMGym: Benchmark Data for Dose-Efficient Autonomous STEM Microscopy
19
 
20
+ This dataset accompanies the [STEMGym](https://anonymous.4open.science/r/STEMGym) benchmark — a Gymnasium-based environment for evaluating autonomous dose-efficient scanning transmission electron microscopy (STEM) agents.
21
 
22
  ## Dataset Description
23
 
24
  STEMGym provides simulated STEM specimens as HDF5 world files. Each world contains:
25
 
26
  - **Overview image**: Low-magnification survey of the full specimen
27
+ - **Tile grid**: High-resolution STEM-HAADF images (128x128 px tiles, 4px overlap, stride=124) arranged in an 8x8 grid
28
  - **Ground truth annotations**: Atom positions, defect types, and phase maps for scoring
29
  - **Metadata**: Pixel size, accelerating voltage, detector geometry, material parameters
30
 
31
  ### Materials
32
 
33
+ | Material | Zone Axis | FOV | Defect Types | Task |
34
+ |----------|-----------|-----|-------------|------|
35
+ | SrTiO3 | [001] | ~100 nm | O vacancies clustered along grain boundary | Defect Census |
36
+ | BaTiO3 | [001] | ~100 nm | Cubic/tetragonal phase boundaries + O vacancies | Phase Mapping + Defect Census |
37
+ | SiGe | [110] | ~50 nm | Ge substitutions concentrated in one quadrant | Targeted Characterization |
38
+ | GaN | [11-20] | ~80 nm | InGaN quantum-well substitutions | Defect Census |
39
+ | Pt nanoparticles | — | ~60 nm | Pt nanoparticles on amorphous carbon | Particle Census |
40
 
41
  ### Difficulty Levels
42
 
43
+ Each material is provided at three difficulty levels:
44
+
45
+ | Level | Vacancy Rate | Phonon Configs | Dose (e-/A^2) |
46
+ |-------|-------------|----------------|---------------|
47
+ | Easy | 5% | 4 | 1e4 |
48
+ | Medium | 3% | 8 | 5e3 |
49
+ | Hard | 1% | 16 | 1e3 |
50
 
51
  ## Files
52
 
 
54
 
55
  | File | Material | Difficulty | Size |
56
  |------|----------|-----------|------|
57
+ | `test_world.h5` | Synthetic (Gaussian blobs) | — | 9.7 MB |
58
+ | `replay_world.h5` | Synthetic (replay validation) | | 2.5 MB |
59
+ | `srtio3_clustered_easy.h5` | SrTiO3 | Easy | 884 MB |
60
+ | `srtio3_clustered_medium.h5` | SrTiO3 | Medium | 884 MB |
61
+ | `srtio3_clustered_hard.h5` | SrTiO3 | Hard | 885 MB |
62
+ | `batio3_interface_easy.h5` | BaTiO3 | Easy | 911 MB |
63
+ | `batio3_interface_medium.h5` | BaTiO3 | Medium | 911 MB |
64
+ | `batio3_interface_hard.h5` | BaTiO3 | Hard | 911 MB |
65
+ | `sige_gradient_clustered_easy.h5` | SiGe | Easy | 150 MB |
66
+ | `sige_gradient_clustered_medium.h5` | SiGe | Medium | 148 MB |
67
+ | `sige_gradient_clustered_hard.h5` | SiGe | Hard | 147 MB |
68
+ | `gan_easy.h5` | GaN | Easy | 451 MB |
69
+ | `gan_medium.h5` | GaN | Medium | 449 MB |
70
+ | `gan_hard.h5` | GaN | Hard | 448 MB |
71
+ | `pt_nanoparticles_easy.h5` | Pt | Easy | 129 MB |
72
+ | `pt_nanoparticles_medium.h5` | Pt | Medium | 129 MB |
73
+ | `pt_nanoparticles_hard.h5` | Pt | Hard | 129 MB |
74
 
75
  ### Model Checkpoints (`checkpoints/`)
76
 
77
  | File | Model | Description | Size |
78
  |------|-------|-------------|------|
79
+ | `atom_finder.pt` | AtomFinderUNet | Atomic column detection ensemble (3 members) | 88 MB |
80
+ | `defect_classifier.pt` | DefectClassifierCNN | Defect type classification | 1.4 MB |
81
+ | `phase_identifier.pt` | PhaseIdentifierResNet | Material phase identification | 7.4 MB |
82
+ | `dqn_agent.zip` | DQN (SB3) | RL navigation baseline | 0.8 MB |
83
+ | `ppo_agent.zip` | PPO (SB3) | RL navigation baseline | 1.0 MB |
84
+ | `sac_agent.zip` | SAC (SB3) | RL navigation baseline | 8.4 MB |
85
+
86
+ ### Transfer Checkpoints (`checkpoints/transfer/`)
87
+
88
+ Material-specific analyst models trained on individual materials for transfer experiments:
89
+
90
+ | Directory | Contents |
91
+ |-----------|----------|
92
+ | `transfer/srtio3/` | atom_finder.pt, defect_classifier.pt, phase_identifier.pt, training_metadata.json |
93
+ | `transfer/batio3/` | atom_finder.pt, defect_classifier.pt, phase_identifier.pt, training_metadata.json |
94
+ | `transfer/sige/` | atom_finder.pt, defect_classifier.pt, phase_identifier.pt, training_metadata.json |
95
+ | `transfer/gan/` | atom_finder.pt, defect_classifier.pt, phase_identifier.pt, training_metadata.json |
96
 
97
  ## HDF5 World Format
98
 
 
102
  /metadata/
103
  pixel_size_nm # Physical pixel size
104
  tile_size_px # Tile dimensions (128)
105
+ grid_shape # Grid dimensions (rows, cols)
106
+ fov_nm # Field of view in nm
107
+ scenario # Material/scenario name
108
+ difficulty # easy / medium / hard
109
+ /overview # (H, W) low-res overview image
110
+ /tiles/{row}_{col} # (128, 128) float32 normalized HAADF tiles
 
111
  /ground_truth/
112
+ atom_positions # (N, 2) in nm
113
+ atom_types # (N,) int [0=pristine, 1=vacancy, 2=substitution]
114
+ defect_mask # (N,) bool
115
+ defect_types # (N,) str ["pristine"/"vacancy"/"substitution"]
116
+ phase_map # (H, W) int32, optional
117
+ /valid_region # (H, W) bool
118
  ```
119
 
120
  ## Usage
 
124
  pip install -e .
125
 
126
  # Download this data
127
+ python stem_gym/scripts/download_data.py
128
 
129
  # Run a benchmark
130
+ stemgym run --agent raster_equipped --task defect_census --world srtio3_clustered_easy --seeds 3
131
  ```
132
 
133
  ## License
 
136
 
137
  ## Citation
138
 
139
+ Citation will be provided after review.