Correlating inline data with final test outcomes in analog/RF devices

In semiconductor manufacturing, a wealth of wafer-level measurements, generally termed inline data, are collected from various on-die and between-die (kerf) test structures and are used to provide characterization engineers with information on the health of the process. While it is generally believed that these measurements also contain valuable information regarding die performances, the vast amount of inline data collected often thwarts efficient and informative correlation with final test outcomes. In this work, we develop a data mining approach to automatically identify and explore correlations between inline measurements and final test outcomes in analog/RF devices. Significantly, we do not depend on statistical methods in isolation, but incorporate domain expert feedback into our algorithm to identify and remove spurious autocorrelations which are frequently present in semiconductor manufacturing data. We demonstrate our method using data from an analog/RF product manufactured in IBM's 90nm low-power process, on which we successfully identify a set of key inline parameters correlating to module final test (MFT) outcomes.