Annotated digital image correlation displacement fields from fatigue crack growth experiments

We present a curated dataset of planar displacement fields from eight fatigue crack growth experiments obtained via full-field digital image correlation (DIC). The dataset covers multiple aerospace-grade aluminium alloys, specimen geometries, material orientations, and load configurations, providing a diverse experimental basis for data-driven fracture mechanics research. Crack tip locations are consistently annotated using an iterative correction procedure applied to all measurements, and fracture mechanical descriptors like stress-intensity factors are provided as additional labels. The dataset comprises 8,794 unique experimentally observed displacement fields and a total of 70,352 supervised samples generated through standardized interpolation and augmentation. DIC data is provided as uniformly interpolated displacement grids at three standardized resolutions 28 x 28, 64 x 64, and 128 x 128 pixels, each available in three dataset sizes to support scalable use cases ranging from educational applications to high-capacity model development. Accompanying metadata and a Python interface facilitate filtering, loading, and integration into reproducible machine learning and fracture mechanics workflows.

💡 Research Summary

The paper introduces “CrackMNIST,” a publicly available, experimentally derived dataset of planar displacement fields obtained from eight fatigue crack growth (FCG) experiments on aerospace‑grade aluminium alloys (AA2024, AA7475, AA7010). The authors collected 8,794 unique full‑field 3‑D digital image correlation (DIC) measurements using a Zeiss Aramis system, and expanded them through systematic augmentation to a total of 70,352 supervised samples. Each experiment varied material condition (rolled or forged), specimen geometry (MT160, CT75), thickness (2–12 mm), crystallographic orientation (LT, TL, SL45), and load ratio (R = 0.1, 0.3, 0.5) while keeping the maximum load constant and applying sinusoidal loading at 20 Hz, following ASTM E647‑15. The DIC acquisition was synchronized with a direct‑current potential drop (DCPD) system, capturing displacement fields at three load levels (minimum, intermediate, maximum) after every 0.5 mm crack‑length increment.

Crack tip annotation follows a three‑stage workflow implemented in the CrackPy toolbox. An initial tip estimate is obtained from a dense line‑intercept detector (0.1 mm spacing). This estimate is refined by iteratively fitting the Williams series to the measured displacement field, and by applying symbolic‑regression‑derived correction formulas (Equation 1) until the correction magnitude falls below 5 µm, yielding sub‑pixel tip accuracy (~0.02 mm). The final tip coordinates are stored as binary masks aligned with the original DIC facet grid.

Stress‑intensity factors (K_I, K_II) and the non‑singular T‑stress are computed directly from the fitted Williams coefficients (Equation 2) for each data point, providing physically meaningful regression targets alongside the displacement fields. Comprehensive metadata accompany every sample, including experiment ID, material, specimen type, thickness, orientation, load ratio, side (left/right), and the instantaneous applied force, enabling flexible filtering and task‑specific dataset construction.

All raw DIC data are interpolated onto uniform square grids of three resolutions—28 × 28, 64 × 64, and 128 × 128 pixels—using bilinear mapping in the physical coordinate frame. The field of view corresponds to roughly 40 mm × 40 mm for MT specimens and 60 mm × 60 mm for CT specimens. The low‑resolution version serves as a lightweight, MNIST‑style dataset suitable for education and rapid prototyping, while the higher resolutions preserve fine displacement gradients essential for high‑fidelity crack‑tip detection and stress‑field regression. For each resolution, three dataset sizes (S, M, L) are provided, allowing users to match computational resources to their needs.

Data augmentation (rotations, flips, added noise) expands the original set, resulting in the 70,352 labeled samples. The authors distribute the dataset via a Python package with a simple API (crackmnist.load_dataset()), which returns the interpolated displacement fields, tip masks, SIF values, and associated metadata. This design promotes reproducible machine‑learning pipelines and lowers the barrier for researchers to benchmark algorithms on real experimental data.

The authors argue that existing fracture‑mechanics datasets are largely synthetic, lacking the noise, resolution limits, and boundary‑condition irregularities inherent to real measurements. By providing a curated, annotated, and openly accessible experimental dataset, CrackMNIST fills a critical gap, enabling systematic benchmarking of supervised learning tasks such as crack‑tip localization, direct SIF prediction, resolution‑dependence studies, data‑efficiency analyses, and uncertainty quantification. The dataset also offers an excellent teaching resource for introducing modern computer‑vision and data‑driven fracture‑mechanics concepts to students.

In summary, CrackMNIST delivers a rich, multi‑modal resource—raw DIC displacement fields, high‑precision crack‑tip annotations, stress‑intensity factor labels, extensive metadata, and a user‑friendly Python interface—positioning it as a foundational benchmark for both academic research and educational applications in data‑driven fracture mechanics.