Knowledge-based software system for fast yield loss detection in a semiconductor fab

ABSTRACT The comparative analysis of process machines in terms of yield related metrics (such as probe and E-Testdata, process and particle data,. ..) is a source of a great deal of information for yield improvement. Withthis aim we published on SPIE's Microelectronic Manufacturing an Advanced Software System to detectmachine-related yield limitors using a comparative analysis. This paper presents the natural expansion ofthat Software System by converting it into a more knowledge-based tool for fast yield loss detection on asemiconductor fab. The new System performs, in an automatic mode, the comparison among machines for every single step selected in the fabrication routing. The detection of statistically significative differences among machines at every step is performed using algorithms that incorporate the overallanalysts experience on our fab. The output of the System allows a fast detection and reaction to yieldissues, mainly to those that are still on the initial or baseline stages.1. INTRODUCTIONThe yield learning and improvement tasks are becoming key issues on nowadays semiconductorfabs [1 1. Time to detection, corrective action implementation and control of process are terms dailyhandled on every yield and failure analysis group. If one adds to this the large size of fabs, large numberof process machines and increasing volume of data retrieved, the global start-point for the yield analystscan be better understood.Within this scope, a global approach to a Yield Analysis Software System was started throughoutLucent Fabs to provide the best in class solution to our yield improvement needs. Once all the toolsdeveloped in the different fabs were unified (after extracting the best of each one of them), we are movingto the next step. we need to be able to build our broad analysts experience into the body of our yieldAnalysis Software Systems.The lack of data variables used to detect problem in the fabrication environment is wealth:process parameters, particle data, E-Test, probe data,. . . We denote all this variables as yield metrics.According to our experience, the comparative analysis of process machines in terms of yield metrics canbe the source of a considerable yield enhancement. In other words, the yield is substantially improved bymatching the performance of all the machines that perform the same type of process (we denote them as