README.md

---
license: cc-by-4.0
task_categories:
  - image-to-image
  - image-classification
tags:
  - microscopy
  - electron-microscopy
  - HAADF-STEM
  - simulation
  - materials-science
  - crystallography
  - webdataset
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: train
        path: "train-*.tar"
---

# Atomagined Dataset (WebDataset)

Simulated atomic-resolution HAADF-STEM images from the [atomagined](https://github.com/MaterialEyes/atomagined) dataset. Each image is a 256x256 simulation of a unique ICSD crystal structure prototype viewed along a specific crystallographic direction, with various post-processing distortion conditions applied.

## Dataset Description

This is a WebDataset conversion of the atomagined proxy dataset (subset available via GitHub). The full dataset (~67,871 images) is hosted on the [Materials Data Facility](https://doi.org/10.18126/szeq-yde5). This version contains 299 samples from 61 unique structure/orientation combinations across three subsets: general simulations, retrieval targets, and retrieval choices.

Each h5 file in the original data contains multiple distortion variants of the same simulation. These have been flattened so that each (structure, distortion) pair is an individual sample.

## Format

WebDataset format: 1 TAR shard, 76 MB total.

Each sample in the shard contains:
- `{idx:06d}.npy` — 256x256 float32 HAADF-STEM image
- `{idx:06d}.json` — metadata

### Metadata Fields

| Field | Type | Description |
|-------|------|-------------|
| `ref_id` | string | Reference ID: `{ICSD_CollCode}.{uvw_encoding}` |
| `subset` | string | Dataset subset: `general`, `targets`, or `choices` |
| `distortion` | string | Distortion condition (e.g., `raw`, `stem00-clean`, `stem00-blur1`, `stem00-cnts2`) |
| `width` | int | Image width (256) |
| `height` | int | Image height (256) |
| `min_value` | float | Minimum pixel intensity |
| `max_value` | float | Maximum pixel intensity |
| `mean_value` | float | Mean pixel intensity |
| `std_value` | float | Standard deviation of pixel intensity |
| `icsd_id` | string | ICSD collection code |
| `chemical_name` | string | Systematic chemical name |
| `chemical_formula` | string | Structural chemical formula |
| `structure_type` | string | Crystal structure type |
| `space_group` | string | Space group name |
| `space_group_number` | string | International Tables space group number |
| `uvw_orient` | string | Viewing direction [u,v,w] |
| `hkl_orient` | string | Miller indices of the viewing plane |
| `n_atoms_projected` | int | Number of atom columns in the projection |

### Distortion Conditions

| Code | Description |
|------|-------------|
| `raw` | Raw simulation output |
| `stem00-clean` | Clean (no distortion) |
| `stem00-blur{1,2}` | Gaussian blur (two levels) |
| `stem00-cnts{1,2}` | Count noise / Poisson noise |
| `stem00-scan{1,2}` | Scan distortion |
| `stem00-bkgd{1,2}` | Background noise |
| `stem00-dfct{1,2}` | Structural defects |
| `stem00-blbk{1,2}` | Blur + background combined |
| `stem01-comb1` | Combined distortion level 1 |
| `stem02-comb2` | Combined distortion level 2 |

## Usage

```python
import webdataset as wds
import numpy as np
from torch.utils.data import DataLoader

url = "path/to/atomagined_wds/train-0000.tar"
dataset = wds.WebDataset(url).decode()
dataloader = DataLoader(dataset, batch_size=32, num_workers=2)

for sample in dataloader:
    image = sample[".npy"]      # 256x256 float32
    metadata = sample[".json"]  # dict with ICSD info, distortion type, etc.
```

## Source

Original data from [MaterialEyes/atomagined](https://github.com/MaterialEyes/atomagined). Full dataset on [Materials Data Facility](https://doi.org/10.18126/szeq-yde5).

## Citation

Schwenker, E. et al. *Image Matching for Computer Vision in Atomic-Resolution Electron Microscopy*, 2020.

## License

CC-BY-4.0