PPM-accuracy Error Estimates for Low-Cost Analog Test: A Case Study

Over the past decade, interest has increased in leveraging low-cost "alternate" test methods as a drop-in replacement for testing analog and RF devices. The current practice for testing such devices, known as specification test, can be very expensive by comparison. By substituting less-costly alternate tests for specification test, substantial test cost reduction can be achieved. Despite this promised cost reduction, the testing community has been reluctant to adopt such alternate test methods in industrial settings, as the error incurred by alternate test can be quite difficult to capture. Estimating the expected test error metrics early in the fabrication process is, therefore, extremely desirable. In this work, we present a case study in which we investigate the reliability of kernel density estimation as a means of providing such early estimates of alternate test error. We also introduce the novel application of Laplacian score feature selection to identify key subsets of alternate tests. With these tools we are able to dramatically reduce the number of alternate tests applied, as well as predict test escape and yield loss error metrics within about 0.5% of their true values when evaluated on an industrial dataset with more than 1 million devices.