Improved alternate test accuracy using weighted training sets

The alternate test paradigm has been proposed as a low-cost replacement to expensive and time consuming conventional specification tests of analog/radio-frequency (RF) integrated circuits. Feasibility of alternate tests may be compromised if the pertinent models that are used for the prediction of a circuit's performance are of poor accuracy. To construct accurate models across the whole design space, a large set of real data needs ideally to be collected from different wafers and lots over a long period of time, that is not possible in the early production stage. In this paper we show that, even when small training sets are available, alternate test accuracy can be significantly improved if the training set is statistically treated in a weighted fashion. The proposed technique is demonstrated for an alternate test procedure developed for a 180nm CMOS RF mixer.