The Microelectronic products from the semiconductor industry are crucial for the many recent technological advances in automotive industries or intelligent production systems but have reliability issues that require failure analysis to investigate the failures' root cause. Moreover, manufacturing is the most significant cost driver in the semiconductor industry, accounting for nearly 40 % of the total cost in the value chain activities. Using the visual inspection system such as cameras or microscopes, images of the microelectronic products or components are acquired to be evaluated for potential defects arising from the manufacturing process.

Using the infra-red (IR) microscopy techniques, images are acquired to detect the chipping defects that may occur during the dicing process. Manual evaluation of such a large amount of data risks subjective errors and backlogs that can drive costs.

In this paper, we developed a defect detection model using advanced machine learning approaches that are more reliable and highly accurate than manual approaches. The task comes with the challenge that chipping defects appear in different shapes and sizes, with some artifacts having similar image features as defects. However, the developed model is based on the supervised learning approach, which includes human expert knowledge in the form of labels to train the model to accurately learn the features that can differentiate actual defects from artifacts. The developed model uses a feature pyramid network (FPN) to detect and localize the chipping defects and artifacts with a better trade-off between accuracy and inspection time when compared to manual inspection.

The trained model was tested on the test set of 199 images, and the model detected defects in the 79 images that were later evaluated by the human subject expert for cross-validation. The cross-validation suggests that 60 images have actual defects, and the rest 19 images have defect-like artifacts that need further analysis. Therefore, applying a deep learning-based approach for failure analysis would assist the experts in overcoming reliability issues and subsequently improve value chain activities.