Published February 12, 2026 | Version v1.1
Dataset Open

Dataset for KBr-KCl replacement using operando synchrotron tomography experiments- part 2:Generative AI

  • 1. ROR icon Utrecht University
  • 1. ROR icon Utrecht University
  • 2. The University of Edinburgh School of GeoSciences
  • 3. ROR icon Paul Scherrer Institute
  • 4. ROR icon RWTH Aachen University
  • 5. ROR icon Universum Bremen

Description

This dataset contains training images and pre-trained models for editing reaction-induced pores from KBr–KCl replacement experiments using the latent space of StyleGAN2-ada. 

The data include images, models and files for images of resolution 512 and 256:

  • TrainingImgsStyleGAN2-ADA.zip : 29k images of size 512 by 512 from KBr-KCl synchrotron experiments in four parts. The whole 4D dataset are available at previous Zenodo repository.

  • TrainingImgsStyleGAN2-ADA_256.zip: Images of size 256 by 256 of KBr-KCl experiments and SEM images of our natural sample.   

Similar data and models for resolution of 256 and 512 can be found in res_256.zip and res_512.zip. These are:

  • Labelled_images: a folder containing the labelled image which are inverted to the latent space of the pre-trained generator.

  • TrainedModels: a folder where you can find pre-trained generator, encoder, and VGG model. 
  • InversionLog (log.txt): a text file logging the loss functions and metrics during inverting images. 

  • labels (df_labels.csv): a csv files showing the labels corresponding to each image in the labelled_images folder.
  • Latent codes (latent_codes.pkl): a pickle file containing the latent codes corresponding to each image in the labelled_images folder.

All datasets were acquired at the TOMCAT beamline, Swiss Light Source (PSI, Switzerland). For more details on the original 4D dataset, see our other Zenodo page for the first part of the paper. You can also read the preprint.

Version 1.1 includes Large 2D SEM images (raw and segmented images) from two orthogonal planes obtained from a monzonite sample showing feldspar replacement. These images have been used to reconstruct 3D images using SliceGAN model as described in the manuscript. These new images are:

  • SEM_Plane1_Raw.tif : Greyscale 2D SEM image of Plane 1 with pixel size of 100 nm.
  • SEM_Plane1_Segmented.tif : Segmented image of Plane 1.  Pore have a pixel value of, the rest is 0.
  • SEM_Plane2_Raw.tif : Greyscale 2D SEM image of Plane 2 with pixel size of 50 nm.
  • SEM_Plane2_Segmented.tif : Segmented image of Plane 2.  Pore have a pixel value of, the rest is 0.

These images were acquired using a Zeiss Gemini SEM.

Files

SEM_Plane1_Raw.tif

Files (2.3 GB)

Name Size Download all
md5:f04f6a2a863a8e02e0cac8700ac35b41
165.5 MB Preview Download
md5:33d2b395960da5079c8614548969a553
132.6 MB Preview Download
md5:1b25cf65c8038a70dc0fa8a8f38e14dc
981.7 MB Preview Download
md5:d77c09b28e7b513d40edf1db8ea1846a
981.7 MB Preview Download

Additional details

Related works

Funding

European Research Council
852069

Dates

Issued
2025-08-29

Software

Repository URL
https://github.com/hamediut/PoreEditGAN
Programming language
Python
Development Status
Active