CrackMNIST: Physics-Guided AI-driven Crack Annotations for Aerospace Materials

Kurzfassung

Introduction Accurate crack tip localization is essential for experimental fracture mechanics based on full-field digital image correlation (DIC) data. Although deep-learning-based methods achieve high accuracy in crack detection, the limited amount of labelled experimental data limits their generalization capabilities and hinders deployment in safety-critical aerospace applications. Here, we present a physics-guided symbolic regression framework that links analytical fracture mechanics, interpretable machine learning, and experimental data analysis and utilize this approach to annotate a large-scale crack dataset obtained by digital image correlation during fatigue crack growth experiments of aerospace-grade aluminum alloys.

Methods Using simulated displacement fields from linear-elastic finite element models under mode I, mode II, and mixed-mode loading, we apply physical deep symbolic regression (1) to discover closed-form crack tip correction formulas expressed in terms of Williams-series coefficients (2). Enforcing physical unit constraints ensures dimensional consistency, reduces the search space, and promotes compact, interpretable analytical expressions.

Results & Discussion The discovered formulas generalize classical correction schemes, converge reliably under iterative application, and recover known theoretical results while extending them to more general loading conditions. We further show how these correction formulas enable a fully automated crack tip annotation pipeline for experimental DIC data (3). Applied to large-scale fatigue crack growth experiments on aerospace-grade aluminum alloys, this pipeline is used to generate a curated benchmark dataset of experimentally measured displacement fields with consistently annotated crack tip locations and fracture-mechanical descriptors. Building on this, the CrackMNIST dataset (4) (available at the public repository (5)) represents the next step toward improving crack tip detection and enabling fast stress intensity factor (SIF) prediction methods. In particular, it provides a standardized experimental benchmark that supports the development and validation of crack tip detection algorithms, stress intensity factor (SIF) estimation methods, and data-driven fracture mechanics models relevant for mechanical testing in the context of physical and virtual certification frameworks. The CrackMNIST dataset offers multiple spatial resolutions and dataset scales, enabling reproducible benchmarking of analytical and data-driven approaches on curated experimental datasets. Furthermore, the methodology is readily extensible to additional materials, loading conditions, and experimental configurations.

Conclusion Overall, this work demonstrates how symbolic regression combined with large amounts of high-fidelity mechanical data can act as an enabling technology in materials science and engineering by discovering new physical relations and mechanical knowledge, and supporting the creation of high-quality benchmark datasets.