Special University Oral Examination

"On the Modeling & Classification of Wafer Map Failure Patterns"

Karen Huyser
Dept of Electrical Engineering

2:15pm, Tuesday May 7, 2002, in Packard 204
Refreshments at 2:00

The talk is open to everyone.


In semiconductor device fabrication, the first phase of quality assurance is conducting process tests during and after wafer manufacture. Unfortunately, process tests qualify wafers, not devices, and many serious device problems cannot be eliminated by the process tests. Instead, the worst device problems are caught during the second phase, called wafersort, when each device on the wafer is subjected to a suite of electrical tests.

When a significant number of devices fail wafersort, the bad devices often cluster in distinctive spatial patterns on the wafer. A 2-D array, called a wafermap, can be constructed to represent the wafer's pattern of pass/fail for any or all tests. Yield engineers use wafermaps to aid in their diagnoses because, together, the failure mode (test ID) and the failure pattern are often characteristic of specific manufacturing problems. Unfortunately, the routine use of wafermaps is rare because of the time required to study them. As a result, potentially valuable information is lost.

The goal of this research has been to make wafermaps easier to use by working toward the automatic classification of wafermap failure patterns. An informal survey at an industrial site was used to discover the most common failure patterns. Particular shapes were selected and formalized mathematically. The talk will identify the shapes that have been modelled and will include a general description of the ideas that led to a statistical formulation. This statistical model of wafermap populations enabled the creation of a computer program to generate arbitrarily large datasets full of synthetic wafermaps.

Four thousand wafermaps, collected at a second industrial site, were re-formatted and classified to create one dataset. Synthetic wafermaps were generated to create additional datasets. Several nearest neighbor classifiers and some prototype classifiers were constructed and used to conduct a variety of experiments. The experiments will be described along with the results. Finally, some recommendations will be offered to yield engineers who might wish to try these methods.