Automated Quantitative Analyses of Fatigue-Induced Surface Damage by Deep Learning

Area of Science:

• Materials Science
• Mechanical Engineering
• Computational Science

Background:

• Digitizing materials is key for product development, but requires understanding microstructure-driven fatigue damage for reliability.
• Surface damage accumulation at the microstructural scale significantly impacts high-performance material lifetime.
• Accurate modeling of fatigue damage is hindered by a lack of comprehensive mechanistic understanding and validated experimental data.

Purpose of the Study:

• To evaluate deep learning (DL) methodologies for semantic segmentation of surface damage in materials.
• To develop and assess image processing approaches for quantitative slip trace characterization.
• To achieve material domain generalizability for DL models in fatigue damage analysis.

Main Methods:

• Utilized a U-Net architecture for semantic segmentation due to limited annotated data.
• Prepared three datasets of SEM images from ferritic steel, martensitic steel, and copper specimens.
• Developed a customized data augmentation pipeline to handle material-specific damage morphology and imaging variance.

Main Results:

• The DL methodology demonstrated material domain generalizability, successfully tested on ferritic steel and conjunct material trained models.
• Multiple image processing routines were implemented to detect slip trace orientation (STO) from DL-segmented extrusion areas.
• The developed DL approach effectively segmented surface damage across different material types.

Conclusions:

• Generalization to multiple materials was achieved for the DL methodology in fatigue damage analysis.
• The study validates the potential of DL for automated analysis of microstructural fatigue damage.
• The findings suggest broader applicability of this DL approach beyond fatigue damage characterization.