Abstract: In semiconductor manufacturing, a low defect rate of manufactured integrated circuits is crucial. To minimize outgoing device defectivity, thousands of electrical tests are run, measuring tens of thousands of parameters, with die that are outside of specified parameters considered as fails. However, conventional test techniques often fall short of guaranteeing acceptable quality levels. Given the large number of electrical tests, it can be difficult to determine which electrical test to rely upon for die quality screening. To address these issues, semiconductor companies have recently begun leveraging artificial intelligence and machine learning to better identify defective devices while minimizing the fallout of good die from electrical tests. To implement these advanced machine learning applications, a novel remote inference capability is also proposed. By placing an inference engine and corresponding machine learning models at the assembly and test house, inferences can be made without any sensitive data leaving the assembly and test house. The result is faster turnaround times on inferences, reduced data loss, increased security, and the enablement of advanced machine learning capabilities for real-time solutions such as adaptive testing.
Keywords: data mining, artificial intelligence