Crack Surface Analysis of Elastomers Using Transfer Learning.

Area of Science:

• Materials Science
• Machine Learning Applications
• Fracture Mechanics

Background:

• Fatigue cracks provide insights into material fracture behavior, including speed, energy dissipation, and stiffness.
• Characterizing fatigue crack surfaces is challenging due to complexity, with existing techniques often being inadequate.
• Machine learning (ML), particularly Convolutional Neural Networks (CNNs), shows promise for image-based material science but requires substantial training data.

Purpose of the Study:

• To develop an effective method for crack surface feature-property mapping using transfer learning (TL).
• To address the data limitations of supervised learning in ML for material science applications.
• To correlate extracted crack features and temperature effects with material properties for predictive modeling.

Main Methods:

• Proposed a modified TL approach by pruning pre-trained CNN models to retain initial convolutional layer weights for feature extraction.
• Employed Principal Component Analysis (PCA) for dimensionality reduction of extracted microstructural features.
• Utilized regression models to correlate crack features and temperature with material properties.

Main Results:

• Successfully applied the proposed method to both artificial microstructures and experimental data of silicone rubbers.
• Established correlations between crack surface features and material properties.
• Developed a predictive model for material property estimation, showing potential to replace extensive experimentation.

Conclusions:

• The developed ML approach effectively extracts relevant features from crack surfaces for material property analysis.
• Transfer learning combined with PCA offers a viable solution for data-scarce scenarios in material characterization.
• The findings suggest a pathway towards computationally driven material property prediction and reduced experimental burden.