Cracks that form during fatigue offer critical information regarding the fracture process of the associated material, such as the crack speed, energy dissipation, and material stiffness. Characterization of the surfaces formed after these cracks have propagated through the material can provide important information complementary to other in-depth analyses. However, because of the complex nature of these cracks, their characterization is difficult, and most of the established characterization techniques are inadequate. Recently, Machine Learning techniques are being applied to image-based material science problems in predicting structure-property relations. Convolutional neural networks (CNNs) have proven their capacity on modeling complex and diverse images. The downside of CNNs for supervised learning is that that they require large amounts of training data. One work-around is using a pre-trained model, i.e., transfer learning (TL). However, TL models cannot be used directly without modification. In this paper, to use TL for crack surface feature-property mapping, we propose to prune the pre-trained model to retain the weights of the first several convolutional layers. Those layers are then used to extract relevant underlying features from the microstructural images. Next, principal component analysis (PCA) is used to further reduce the feature dimension. Finally, the extracted crack features together with the temperature effect are correlated with the properties of interest using regression models. The proposed approach is first tested on artificial microstructures created by spectral density function reconstruction. It is then applied to experimental data of silicone rubbers. With the experimental data, two analyses are performed: (i) analysis of the correlation of the crack surface feature and material property and (ii) predictive model for property estimation, whereby the experiments can be potentially replaced altogether.
Keywords: convolutional neural network; elastomers; fatigue; machine learning; silicone; spectral density function; transfer learning.