Kernel Smoothing Technique Based on Multiple-Coordinate System for Screening Potential Failures in NAND Flash Memory

With the growing complexity of integrated circuits and increasing capacity of memory devices, it is becoming increasingly difficult to screen out all defective dies through wafer test. To address this problem, statistical methods are being increasingly applied on test data to screen out potential defective dies. Among such methods, the good-die-in-bad-neighborhood (GDBN) method is effective. Most GDBN methods use the bin data of nearby dies through Cartesian coordinates. However, the bin data provide only binary information, and Cartesian coordinates do not fully reflect the spatial correlations of wafers. Therefore, this paper proposes a novel GDBN method using bad block data, which are key indicators of the in-field failure of NAND flash memory. The defect risk is predicted through statistical kernel smoothing based on Cartesian and polar coordinates to consider the spatial correlations in various environments of wafer fabrication. Experiments on NAND flash products show that the in-field failure classification performance of the proposed method, in terms of the area under the curve value, is 30% higher than that of the existing method, based on the bin data and Cartesian coordinates.