Fuzzy logic automates semiconductor-wafer inspection
To assess problems during semiconductor manufacturing, test engineers use both high-resolution images from individual wafers and low-resolution-defect wafermap images from in-line inspection tools. By analyzing the spatial distributions or "signatures" of the common defects on wafer surfaces, potential problems in the manufacturing process can be identified, classified, and resolved.
To automate this inspection process, Kenneth Tobin, group leader of the Image Science and Machine Vision Group at Oak Ridge National Laboratory (ORNL; Oak Ridge, TN), and his colleagues Shaun S. Gleason and Thomas P. Karnowski have developed spatial-signature-analysis (SSA) software that performs clustering and classification of wafer-defect signatures.
Developed in a cooperative research and development agreement between ORNL and Sematech (Austin, TX), the software, known as SSAtool, segments defect data into high-level groupings and classifies the defect signatures into user-defined categories using fuzzy-logic techniques. Although Sematech retains the right to license the software in semiconductor manufacturing, the ORNL can license it in other application areas.
"To date, the SSA technology has been very successful," says Tobin. "There are currently six licenses given to semiconductor suppliers and seven to semiconductor manufacturers," he adds. Originally developed under the Sun Microsystems` Solaris-UNIX environments, the software has also been ported to Microsoft`s Windows NT/95, IBM`s AIX UNIX, Hewlett-Packard`s UNIX, and Digital Equipment`s Alpha-based systems.
In operation, the software first measures the features of the objects found on the wafer and then groups them as either global, amorphous, curvilinear, or microcluster. After this grouping, feature measurements are made that relate a particular spatial signature with a specific manufacturing problem. This higher-level classifier is trained by the user, who next adds the examples of various process signatures to a training database. Once a training set exists, the software auto matically segments and classifies multiple signatures from a standard electronic wafer map file into user-defined categories.
"To automatically label process signatures based on user definitions, a fuzzy pair-wise classification algorithm is used that relies on statistical, morphological, and inspection measurements of objects in the wafer image," says Tobin. The classifier contains an algorithm for automatically ranking and selecting the features necessary to distinguish among defined classes.
"This automatic feature-selection procedure requires no input from the user and results in a system that is simple to use and maintain," states Tobin.
To store, retrieve, edit, and manage signature examples, Tobin, Gleason, and Karnowski have also developed a Prototype Signature Library (PSL). Designed to provide an intuitive display of alphanumeric and color codes in conjunction with thumbnail images of the wafermap classes to the user, the PSL also highlights any possible conflicts and ambiguities that exist in the training data.
SSA automates the analysis of wafermaps, which isolate the semiconductor-manufacturing problems based on common wafer signatures. Better yet, the generation of detailed SSA descriptions provide new information for statistical-process-control systems and can reduce the number of wafers required for off-line review using optical or scanning-electron-microscope methods.