Material defects remain one of the primary challenges in the production of compound semiconductor devices. Substrates used for manufacturing integrated circuits (ICs) often exhibit high defect densities, which can severely impact downstream yield during wafer processing and die assembly. Understanding the complex relationship between material defects and final device yield is far from straightforward, due to the multitude of defect types and their interactions during fabrication.
In this work, we present an end-to-end yield management platform that has been successfully deployed across more than ten manufacturers of compound semiconductor devices. This platform integrates data from multiple stages of the production process, ranging from material inspection to electrical wafer sort, providing a comprehensive dataset that spans the entire manufacturing pipeline. End-to-end data allows building advanced artificial intelligence and machine learning (AI/ML) models capable of predicting defect-limited yield based on substrate defects.
The predictive models address several key challenges in yield management. First, we calculate the “kill ratio,” which identifies strong killer defects—specific defect types whose presence alone is sufficient to cause die failure. Second, we perform a correlation analysis to filter out defect types whose increasing counts are paradoxically correlated with passing die, allowing us to focus only on meaningful defect predictors.
After this filtering stage, we apply the classification ML model to capture complex, nonlinear relationships between defect patterns and die outcomes. Importantly, the platform allows users to set flexible probability thresholds within the model to classify die as passing or failing. This feature gives users the ability to adjust the balance between underkill (failing to catch defective die) and overkill (rejecting good die) depending on their specific revenue priorities, product mix, and risk tolerance.
We benchmarked the approach using Silicon Carbide (SiC) use cases, which are representative of wide bandgap compound semiconductors used in power electronics and high-frequency applications. The models demonstrated the ability to accurately identify killer defects and facilitate development of the rules for substrate grading.
Overall, this integrated yield management platform not only improves yield prediction accuracy but also enhances root cause analysis, process control, and supplier feedback loops. It represents a critical step forward in bridging the gap between material quality and device performance in compound semiconductor manufacturing.
