Adaptive Test for Mixed-Signal Chips Based on Information Gain and Backpropagation Neural Network

Mixed-signal chips, serving as critical interfaces between analog and digital domains, are playing an increasingly important role in modern electronic systems. To accelerate manufacturing tests, adaptive testing has emerged as a promising approach that dynamically adjusts the test sequence, prunes redundant items, and detects defects at minimal cost. Existing studies predict the performance of devices under test (DUTs) from low-cost test items in current and prior stages with different predictors. However, a systematic analysis for selecting appropriate prediction pairs is still lacking. Consequently, a large number of DUTs are required for training, which increases the overall test time and reduces efficiency. To overcome the limitations, this study introduces a correlation-driven method for test-item selection, enabling accurate DUT classification in adaptive testing. Test items are recategorized into “leader,” “follower,” and “uncorrelated” groups based on their extra information gain cost (EIGC), and a backpropagation neural network (BPNN) regression model is used to predict the performance of the “follower” set. The nondominated sorting genetic algorithm is implemented for threshold setting and test-item categorization, considering both test time cost and accuracy. Experimental validation on a commercial mixed-signal chip demonstrates the effectiveness of the proposed method, achieving a defect escape rate as low as 1% with a 40% reduction in test time.