edfas.org ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 26 NO. 3 6 APPLIED AND TESTED DEEP LEARNING MODELS AND EVALUATION METRICS These experiments tested four different U-Net[30] architectures: U-Net 2D, U-Net++,[31] U-Net 3D,[32] and a modified U-Net 2D. In the modified U-Net 2D, a drop-out layer was added after each convolutional block of U-Net’s original model to avoid overfitting on limited training data. In addition to selecting the architecture, the impact of different loss functions on model training is also investigated. Categorical cross-entropy (CCE), Dice loss, and Jaccard loss are tested with all model architectures.[33] This comparative approach allows identifying the most suitable model and combination of loss functions for accurately detecting voids within solder balls. A third-party software, Dragonfly,[29] is used for deep learning model training. This software offers functionalities to handle and visualize 3D data. The AI modules within the software provide a no-code-based solution for different deep learning model training. To ensure consistency and fair performance comparison across the different U-Net models, all models were trained from scratch within an identical training environment. The Adadelta[34] optimizer was used to optimize the model parameters during the training process. Early stopping and reduced learning rate on plateau techniques were implemented during training to prevent overfitting and enhance generalization performance. (a) (b) Fig. 2 Region of interest visualization of 3D XRM data of solder balls and its label. (a) 3D visualization of the region of interest extracted from raw XRM data, having nine solder balls. (b) Corresponding label in 3D (3D data sliced to the half for better visualization of voids inside solder balls), where green indicates the solder balls and red indicates the voids. (a) (b) Fig. 3 Region of interest visualization of 2D XRM slice extracted from 3D XRM data of solder balls and its corresponding 2D label. (a) 2D slice (region of interest), and (b) corresponding label in 2D, where green color indicates solder ball (marked in a circle), red color is void in solder balls (marked in a square), and purple color is background.
RkJQdWJsaXNoZXIy MTYyMzk3NQ==