Text

theses

In this dissertation, we present several methodologies for mining data obtained in semiconductor manufacturing processes. We first present a new step-down spatial randomness test aiming at automatically detecting abnormal dynamic random access memory (DRAM) wafers with multiple spatial maps. Testing the spatial randomness of a DRAM wafer is challenging. A DRAM wafer includes multiple spatial maps, resulting to a more complex and lengthy testing process compared with that of a single wafer map of a flash memory. To monitor the spatial randomness of the multiple spatial maps, we propose a new step-down spatial randomness test to detect abnormal DRAM wafers. In the proposed methodology, we adopt nonparametric Gaussian kernel-density estimation to transform the original fail bit test (FBT) values into binary FBT values. We also propose a spatial local de-noising method to eliminate noisy defect chips to distinguish the random defect patterns from systematic ones. Secondly, we propose a novel matrix factorization method, called regularized singular value decomposition (RSVD), which aims at the automated classification of chip level failure patterns on fail bit map (FBM) of each DRAM chip. The RSVD based approach decomposes a FBM into several binary eigen-images to extract features that can provide the characteristics of the failure patterns on the FBM. By employing the extracted features as input vectors, k-nearest neighbor (k-NN) classifier is applied to classify feature patterns on a FBM into either single bit failed one or non-single bit failed one. Finally, we propose a new Bayesian classification model for uncertain data to classify abnormal DRAM wafers that include spatial features with uncertainty. Bayesian classifier has been extensively used for the classification of certain data. However, since every data object in the uncertain data is not represented by a point value, it is difficult to directly apply the Bayesian classifier for certain data. In the proposed approach, the multivariate kernel density estimate for uncertain data is proposed to estimate the class conditional probability density function (pdf). We then apply the Bayes theorem to calculate the posterior probability of a testing data object based on the estimated class conditional pdf.

ETD_6257

Ph.D.

Includes bibliographical references

by Byunghoon Kim

doi:10.7282/T3FQ9ZF6

ETD doctoral

The author owns the copyright to this work.