Explainable machine learning for precise fatigue crack tip detection.

Area of Science:

• Materials Science
• Computer Vision
• Mechanical Engineering

Background:

• Deep learning models offer breakthroughs in computer vision but lack domain knowledge, hindering acceptance in critical applications.
• Explainability is vital for deploying deep learning in safety-critical fields like aircraft component analysis.

Purpose of the Study:

• To develop and evaluate a novel deep learning architecture, ParallelNets, for crack tip detection in fatigue crack growth experiments.
• To compare the performance and explainability of ParallelNets against a U-Net architecture.

Main Methods:

• Training convolutional neural networks (CNNs) on full-field displacement data from digital image correlation (DIC).
• Introducing ParallelNets, a hybrid architecture combining segmentation and regression for crack tip localization.
• Utilizing Grad-CAM for visualizing neural attention and assessing model interpretability.

Main Results:

• ParallelNets demonstrated superior accuracy, robustness, and stability compared to the U-Net architecture.
• Grad-CAM visualizations confirmed ParallelNets' focus on physically relevant crack tip regions.
• The proposed architecture enhances the understanding and acceptance of deep learning in materials science.

Conclusions:

• ParallelNets offers a more explainable and performant solution for crack tip detection using deep learning.
• Integrating domain knowledge through architecture design improves model reliability in fatigue crack growth studies.
• Explainable AI methods are crucial for validating deep learning applications in engineering and materials science.