In the microelectronics industry, crack inspection of bond pads is a critical quality control step, often hindered by the large volume of samples requiring analysis. This process is typically manual and time-consuming, placing a significant burden on engineers tasked with visually inspecting each sample for defects. While some automated approaches have been developed, they are generally limited to specific sample types, reducing their applicability in diverse manufacturing environments.

In this study, we developed a semi-automated imaging and analysis pipeline for the robust localization of bond pads on a wide variety of microelectronic chips. The workflow is implemented as a ZEN core job template, enabling seamless integration within the ZEISS ZEN core microscopy software environment. It incorporates a deep learning-based object detection model capable of operating on both pre-acquired and newly captured datasets. The model outputs precise spatial coordinates of each detected bond pad, which are compiled into a positions list to facilitate targeted re-imaging following sample modification or processing.

The key challenges considered in our approach were as follows:

• Model fit well for one type of chip only
• Samples are very diverse in terms of internal structure, geometry, color and size. The visual appearance of bond pads can differ markedly due to variations in shape; and the presence or absence of bonded wires
• Reproducibility for automation here is challenging

When used in conjunction with a compatible sample holder that ensures repeatable sample positioning, the workflow becomes fully automated, eliminating the need for manual intervention during reacquisition.

In order to overcome the challenges, we trained a single-shot object detection model that generalizes on more than 25 sample types, is 30x faster than conventional image processing approaches, including manual detection, and achieved a 0.98 precision and mAP0.5-0.95 of 0.85.