HMT-SEM silver-particle images, microstructural line annotations and frozen analysis data
Description
This dataset contains 48 original HMT-SEM images, corresponding manual microstructural line annotations, fixed image-level train/validation/held-out splits (34/7/7), physical-scale metadata, and frozen intermediate and classification results supporting the associated manuscript. Four independent classifiers associate segmented silver particles with grain boundaries and screen slip-band candidates. Labels derived from manual lines are weak supervision and an evaluation reference, not independently assigned particle ground truth. The package includes source data for the current figures, archived MATLAB models/code, and an integrity/metric verification script. Training from scratch has not been independently reproduced from this package; historical wide-field training image caches were not recovered.
Files
Steps to reproduce
1. Download the complete dataset and preserve its directory structure. Read README.md and use 03_metadata/image_index.csv to locate each SEM image and its annotations, and to identify the fixed split of 34 training, 7 validation, and 7 held-out images. 2. From the directory containing MANIFEST.csv, run python verify_release.py with Python 3.9 or later. The script checks file integrity and the image split, then averages the saved per-image precision, recall, and F1 scores over eligible images and compares the results with the archived summary tables. A successful run reports "status": "passed". 3. To recreate the plots in Figures 3, 6, 7, and 8, use the CSV files in 07_figure_source_data and follow the README in that directory. 4. For the archived MATLAB code and models, see 08_code/README.md for the software requirements and directory layout. End-to-end retraining from the original images remains unverified.
Institutions
- Nanjing Tech UniversityJiangsu, Nanjing