Stemson-AI/nanoparticle-size-estimation

Nanoparticle Size Estimation Dataset (WebDataset)

100,000 simulated HAADF-STEM images of nanoparticles with three image variants per sample (raw simulation, normalized, and CycleGAN style-transferred), paired with per-structure metadata including atomic coordinates, morphology, and simulation parameters.

Dataset Description

Synthetic HAADF-STEM images of Pt nanoparticles on various support interfaces, generated for nanoparticle size and morphology prediction. Each sample includes the raw simulation output, a set-normalized version, and a CycleGAN-transferred version that emulates experimental noise characteristics. Associated with the paper Nanoparticle size estimation.

Original data: Zenodo record 14608502

Format

WebDataset format: 20 TAR shards (~1 GB each), 19 GB total.

Each sample in a shard contains:

• {idx:06d}.sim_raw.npy — raw simulated HAADF-STEM image (128x128, float32)
• {idx:06d}.sim_norm.npy — set-normalized simulation (128x128, float32)
• {idx:06d}.cyclegan.npy — CycleGAN style-transferred image (128x128, float32)
• {idx:06d}.json — metadata

Metadata Fields

Field	Type	Description
sample_id	int	Original sample index
original_filename	string	Original .npy filename
width	int	Image width (128)
height	int	Image height (128)
pixel_size	float	Pixel size in Angstroms
beam_offset_x	float	Beam offset in x
beam_offset_y	float	Beam offset in y
particle_rotation	float	Particle rotation angle (radians)
structure_type	string	Nanoparticle morphology (cluster, wulff, random, etc.)
support_interface	string	Support surface orientation (e.g., 111, 100, random)
particle_interface	string	Particle surface orientation
element	string	Atomic species (Pt)
subset	string	Dataset subset label (e.g., AC_111_111)
n_atoms	int	Number of atoms in the structure
raw_min	float	Minimum pixel intensity (raw image)
raw_max	float	Maximum pixel intensity (raw image)
raw_mean	float	Mean pixel intensity (raw image)
raw_std	float	Std dev of pixel intensity (raw image)

Image Properties

• Count: 100,000 samples (300,000 images total across 3 variants)
• Format: NumPy arrays (.npy), float32
• Resolution: 128x128
• Modality: Synthetic HAADF-STEM
• Element: Pt nanoparticles
• Morphologies: cluster, wulff, random

Usage

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

url = "path/to/nanoparticle_wds/train-{0000..0019}.tar"
dataset = wds.WebDataset(url).decode().shuffle(1000)
dataloader = DataLoader(dataset, batch_size=64, num_workers=4)

for sample in dataloader:
    raw_image = sample[".sim_raw.npy"]
    norm_image = sample[".sim_norm.npy"]
    cyclegan_image = sample[".cyclegan.npy"]
    metadata = sample[".json"]

Source

Original data from Zenodo record 14608502. Associated paper: Nano Letters (2024).

License

CC-BY-4.0