1. ABOUT THE DATASET -------------------- Title: Reconstruction of 3D crystal growth from transmission optical microscopy images - Supporting data Creators: Thomas P. Ilett, Thomas A. Hazlehurst, Chen Jiang, Cai Y. Ma, David C. Hogg, Kevin J. Roberts Organisation: University of Leeds Rights-holder: Copyright 2026 University of Leeds Publication Year: 2026 Description: This dataset contains synthetic RGB images of randomised α-LGA crystal structures, including noisy and clean renderings, 3D keypoint and wireframe heatmaps, randomised surface bumpmaps, and ground-truth 3D mesh vertices. It also includes full scene configurations, crystal growth parameters, randomised texture and lighting details, summary statistics, and a predefined training/testing split. Included alongside this are the experimental data (images and manual measurements) and result videos. Cite as: Thomas P. Ilett, Thomas A. Hazlehurst, Chen Jiang, Cai Y. Ma, David C. Hogg, Kevin J. Roberts (2026): Reconstruction of 3D crystal growth from transmission optical microscopy images - Supporting data. University of Leeds. [Dataset]. https://doi.org/10.5518/1684 Related publication: Thomas P Ilett, Thomas A Hazlehurst, Chen Jiang, Cai Y Ma, David C Hogg, Kevin J Roberts, Reconstruction of 3D crystal growth from transmission optical microscopy images, PNAS Nexus, 2026;, pgag080, https://doi.org/10.1093/pnasnexus/pgag080 Contact: K.J.Roberts@leeds.ac.uk 2. TERMS OF USE --------------- This dataset is licensed under a Creative Commons Attribution 4.0 International Licence: https://creativecommons.org/licenses/by/4.0/. 3. PROJECT AND FUNDING INFORMATION ---------------------------------- Title: Advanced Crystal Shape Descriptors for Precision Particulate Design, Characterisation and Processing (Shape4PPD) Dates: 1 October 2021 - 31 December 2024 Funding organisation: Engineering and Physical Sciences Research Council (EPSRC) Grant no.: EP/W003678/1 The authors are grateful for the financial support of EPSRC UK through the Shape4PPD project (EP/W003678/1) in collaboration with AstraZeneca, Cambridge Crystallographic Data Centre, Infineum, Keyence, Pfizer, Roche, Syngenta, Imperial College, the University of Hertfordshire and the University of Strathclyde. One of the authors (CJ) thanks EPSRC UK for the Doctoral Training Partnership award (EP/W524372/1), which is co-sponsored by Syngenta in collaboration with Neil George and Jennifer Webb. Thanks are also due to the EPSRC Centre for Doctoral Training in Complex Particulate Products and Processes (EP/L015285/1) who support CJ as an aligned student. We also acknowledge the funding support in digital design workflow modelling of crystallization processes from the UK's Advanced Manufacturing Supply Chain Initiative through the `Advanced Digital Design of Pharmaceutical Therapeutics' (ADDoPT) project (grant No. 14060) in collaboration with AstraZeneca, Bristol–Myers Squibb, Britest, Cambridge Crystallographic Data Centre, GlaxoSmithKline, Perceptive Engineering, Pfizer, Process Systems Enterprise, STFC’s Hartree Centre, and the Universities of Cambridge and Strathclyde. The project used the Tier 2 HPC facility JADE2, funded by EPSRC (EP/T022205/1). 4. CONTENTS ----------- >> crystal_bumpmaps/ Normal-map-style randomised bumpmaps for each crystal to be applied the surface of the crystal objects as a texture. >> ds_stats.yml Summary statistics for the dataset, including minimums, maximums, means and variances for all of the parameters. >> experiments/ Original microscope images and manual measurements for the nine experiments reported in the paper. >> images/ RGB renderings of synthetic crystal configurations including randomised sources of noise. >> images_clean/ RGB renderings of synthetic crystal configurations without randomised sources of noise. >> keypoints_wfv=1.0_kpv=1.0/ Generated 3D keypoint and wireframe heatmaps stored as three-channel images where the first channel contains Gaussian blobs of variance 1 px at each keypoint location, the second channel contains all of the camera-facing wireframe edges and the third contains all of the refracted wireframe edges. >> keypoints_wfv=1.0_kpv=10.0/ As above, but with larger variance (10 px) used to generate the Gaussian blobs at the keypoint locations. >> options.yml Full configuration settings used for dataset generation and synthetic image rendering. >> parameters.csv Crystal configuration parameters, canonical distances, scale, image references and computed face areas. >> rendering_parameters.json Detailed representation of each synthetic crystal scene including randomised lighting configurations, randomised textures and bumpmap generation parameters and crystal seed configurations. >> train_test_split_0.80.json Dataset sample indices for training and testing split (80/20), used in all experiments. >> validation/ Additional validation images included to verify that the dataset images can be completely reconstructed from the data saved in rendering_parameters.json. >> vertices.json 3D mesh vertex data for each synthetic crystal, used as ground truth for geometric losses. >> videos/ Videos files demonstrating the method and results as described in the supplementary information. 5. METHODS ---------- Please refer to the main paper for details of the crystal shape generation and image rendering methods.