DiffraNet

Dataset description

This repository contains the DiffraNet diffraction image dataset introduced by Souza et al. in the paper "DeepFreak: Learning Crystallography Diffraction Patterns with Automated Machine Learning". The original dataset consists of 512×512 grayscale diffraction images, including both synthetic and real experimental data. The synthetic images were generated using the nanoBragg simulator, while the real images were collected from serial crystallography experiments. :contentReference[oaicite:0]{index=0}

This repository mirrors the original dataset and additionally includes a modified version of the raw real dataset (real_raw_mod) for my own purposes.

Credit

The original DiffraNet dataset was created by:

• Artur Souza
• Leonardo B. Oliveira
• Sabine Hollatz
• Matt Feldman
• Kunle Olukotun
• James M. Holton
• Aina E. Cohen
• Luigi Nardi

Paper

Souza, A. et al. DeepFreak: Learning Crystallography Diffraction Patterns with Automated Machine Learning (2019)

https://arxiv.org/abs/1904.11834

Original project

https://arturluis.github.io/diffranet/

All credit for the original dataset belongs to the authors. This repository is intended only as a mirror of the original dataset and includes one additional modified split (real_raw_mod) described below.

---

Dataset structure

The repository contains four top-level directories:

synthetic/
real_raw/
real_preprocessed/
real_raw_mod/

Each of the directories represents a different dataset or a different dataset version.

The original DiffraNet dataset consisted of synthetic, real_raw, real_preprocessed.

real_raw_mod is a modified version of real_raw for use in transfer learning, and was not a part of the original dataset.

The folder hierarchy defines both the dataset split and the class labels.

synthetic/

Synthetic diffraction images generated with the nanoBragg simulator.

Dataset splits:

• training
• validation
• test

Classes:

1. blank
2. no crystal
3. weak diffraction
4. good diffraction
5. strong diffraction

The original dataset contains approximately 25,000 synthetic images.

real_raw/

Original real diffraction images provided by the DiffraNet authors.

Characteristics:

• cropped to 512×512
• contain the experimental beamstop shadow

Dataset splits:

• validation
• test

Classes:

1. no diffraction
2. diffraction

real_preprocessed/

Original preprocessed version of the real dataset.

The images are identical to real_raw except that pixel intensities have been rescaled so that the mean pixel value matches that of the synthetic dataset.

The dataset organization is identical to real_raw.

real_raw_mod/

Modified version of real_raw prepared for supervised learning.

The original validation split has been subdivided into

• training
• validation

while the original test split has been left unchanged.

---

Dataset statistics

Dataset	Images	Classes
Synthetic	~25,000	5
Real (raw)	457	2
Real (preprocessed)	457	2

All images are 512×512 grayscale.

---

Citation

If you use this dataset, please cite the original paper:

@article{souza2019deepfreak,
  title={DeepFreak: Learning Crystallography Diffraction Patterns with Automated Machine Learning},
  author={Souza, Artur and Oliveira, Leonardo B. and Hollatz, Sabine and Feldman, Matt and Olukotun, Kunle and Holton, James M. and Cohen, Aina E. and Nardi, Luigi},
  journal={arXiv preprint arXiv:1904.11834},
  year={2019}
}