Data and trained model weights released alongside our papers. Everything is hosted on Hugging Face or Zenodo, and is free to download.
HAADF-STEM images with acquisition metadata (CIMP)
7,330 HAADF-STEM images, each paired with the seven acquisition parameters used to record it: pixel size, dwell time, convergence angle, beam current, gain, offset and inner collection angle. Images are stored raw as 16-bit HDF5, mostly 2048 x 2048. Released with our AI4Physics workshop paper at ICML 2026. The files currently live in a public storage bucket rather than a dataset repository, so there is no dataset card yet.
3D morphology prediction from single micrographs
Training data and model weights for predicting the 3D atomic structure of supported nanoparticles from a single STEM image with a point cloud diffusion model. Published in npj Computational Materials, 2026.
In situ tracking of indium oxide catalysts
Data and weights for the U-Net that segments contrast changes in in situ gas-cell TEM time series, used to track In2O3 dynamics during CO2 hydrogenation to methanol. Published in Advanced Materials, 2025.
Multislice surrogate model
Synthetic STEM images and weights for the 3D U-Net that replaces multislice simulation, running 1000x faster at 128 x 128. Published in Nano Letters, 2025.
Nanoparticle sizing with generative AI
CycleGAN style transfer between simulated and experimental STEM images, with a size estimator that predicts the number of atoms in a nanoparticle. Published in Small Methods, 2024.
Atomic column localisation and segmentation
Synthetic training data and weights for locating atomic columns at picometer precision in noisy STEM data, and classifying them as particle or support. Published in npj Computational Materials, 2024.
