Study on pore features in sintered die-attach microstructures based on machine learning

Low-temperature sintering of nano-sized metal particles is a promising approach for die-attach bonding, where the porous microstructure plays a critical role in determining bonding performance. However, the internal relationships among various microstructural features and their effectiveness in characterizing morphology remain unclear. In this study, a series of pore-related features were extracted from massive cross-sectional SEM images and analyzed using machine learning. Correlation analysis revealed mathematical relationships among features, and accordingly they can be categorized into two groups related to pore distribution and shape. Principal component analysis was employed to obtain representative but interpretable descriptors from the correlated features. Four machine learning models, including K-nearest neighbor, support vector machine, random forest, and artificial neural networks, were trained on the transformed dataset, achieving over 90% accuracy in classifying images captured from samples prepared under different conditions. In addition, a variational autoencoder-artificial neural network framework was constructed for comparison, demonstrating the effectiveness of the proposed physical feature extraction. Significance analysis based on the four models reveals the accuracy and pertinence of these descriptors in microstructure-property assessment.