Semiconductor probe card proactive maintenance using graph self-supervised learning and an empirical study

Abstract

Semiconductor probe cards for wafer testing are increasingly challenging owing to the high density of bonding pads on the surface of integrated circuits (IC), the growing complexity of IC device features, and the extensive customization required for various IC products. Early fault detection and resolution are crucial for proactive maintenance to maintain the optimal operational performance and overall equipment effectiveness of wafer probing test equipment. Furthermore, the increasing number of possible corrective solutions for probe cards and the lack of concatenated labels for model training, along with the heterogeneity of domain knowledge for similar problems, has made it increasingly difficult for the engineers to find effective solutions for abnormal symptoms. Most of the existing studies focus on fault diagnosis, failure detection and classification, and advanced equipment control. As part of a continuous effort to fill the gaps, this study aims to develop a UNISON framework for graph self-supervised learning that integrates knowledge graph (KG), graph convolutional neural networks (GCN), and self-supervised learning to predict the conditions and effectively recommend the optimal list of corrective actions for early detection and resolution of the detected abnormal symptoms. To validate the proposed approach, an empirical study was conducted in a leading semiconductor testing company in Taiwan. The results have shown practical viability of the developed solution that can effectively assist the engineers in selecting the corrective actions for proactive maintenance to reduce machine downtime and enhance customer satisfaction.