CrackMNIST - Annotated Digital Image Correlation Displacement Fields from Fatigue Crack Growth Experiments

CrackMNIST is a curated, large-scale, and standardized dataset of experimentally acquired digital image correlation (DIC) displacement fields from fatigue crack growth experiments. The dataset is designed to support data-driven and physics-informed machine learning research in fracture mechanics, with a particular focus on crack tip detection and the prediction of fracture-mechanics descriptors from full-field measurements.

The dataset is derived from eight fatigue crack growth experiments conducted on aerospace-grade aluminium alloys (AA2024, AA7475, AA7010) covering multiple specimen geometries, material orientations, thicknesses, and load ratios. Full-field planar displacement fields were measured using a commercial Zeiss Aramis Stereo DIC system and consistently post-processed across all experiments.

Crack tip locations are provided as high-accuracy annotations, obtained via an iterative crack tip correction procedure based on Williams-series fitting, achieving sub-pixel crack tip accuracy.
In addition to crack tip segmentation masks, CrackMNIST includes fracture-mechanics labels, namely the mode-I and mode-II stress intensity factors and the T-stress.

To ensure scalability and broad usability, the dataset is provided at three standardized spatial resolutions (28 x 28, 64 x 64, and 128 x 128 pixels), each representing the same physical region around the crack tip. For every resolution, three dataset sizes (S, M, L) are released with fixed and reproducible training, validation, and test splits, supporting applications ranging from lightweight educational use to high-capacity deep learning benchmarks.

Overall, CrackMNIST comprises 8,794 unique experimentally observed displacement fields and a total of 70,352 supervised samples generated through standardized interpolation and physically consistent data augmentation.

Each sample includes:

• two-channel planar displacement fields,

• crack tip segmentation masks

• fracture-mechanics descriptors (K_I, K_II, T),

• and rich metadata such as experiment ID, material, specimen type, orientation, load ratio, and applied force.

The dataset can be accessed either by direct download of the HDF5 files or via the accompanying open-source Python package crackmnist, which provides convenient data loading, filtering, and integration into reproducible machine learning workflows.

CrackMNIST aims to serve as a benchmark dataset for experimental fracture mechanics, enabling systematic comparison of learning-based crack tip detection methods, resolution studies, uncertainty quantification, and hybrid physics-ML approaches, while also providing a high-quality resource for teaching and education.

For further information, code, and usage examples, see: https://github.com/dlr-wf/crackmnist