Data mining methodologies for yield prediction in semiconductor manufacturing.

Abstract

Increasing yield is a primary concern to IC manufacturing companies as it dictates the readiness of a new process for high volume manufacturing. In order to expedite the process of discovering yield issues, companies have started looking for ways to perform early prediction for such issues. This thesis evaluates different classification models for classifying wafers during in-line processing. We first consider SVM as an effective model for classifying multivariate, multi-modal, and inseparable data points, by deriving multidimensional hyperplanes to classify wafers into lowand high-yield classes. The thesis includes a proposal for how the model can be applied for yield classification, and how it can be adaptively updated in a manufacturing environment. It shows how to select the SVM parameters values for best yield classification. Furthermore, performance evaluation is conducted on real manufacturing data, comparing the proposed model to state of the art. Results show that, in all cases of conducted experiments, the SVM model consistently outperforms the other models with and without adaptive model updates. It is also shown that the classification model can be built and executed using a reduced set without compromising its accuracy. In addition to SVM evaluation on real data, we also propose to evaluate the different models with simulation under disturbance situations in semiconductor manufacturing. Towards this goal, we develop a simulator that captures shifts in inline measurements, and we study the related yield effects under different combinations of available inline parameters. Over 1120 experiment scenarios are considered in this evaluation. Results provide a critical insight into best choice of classifiers for yield prediction. The experiments indicate that relative classifier performance can change depending on the state of manufacturing data and the parameters causing yield issues. For example, SVM performs better with complex multi-variate situations as demonstrated in the first part of the thesis, whe