Wafer die yield prediction by heuristic methods

Yield is a very important criterion to measure the semiconductor wafer fabrication facilities (FABs) productivity. The finished products will be check by Wafer Acceptance Test (WAT) and Circuit Probe (CP) to classified into ferior goods or inferior goods. This research applied the data from WAT and CP for the selection of the most important measuring parameters to improve the yield. Three methods, namely Support Vector Regression (SVR), Group Method of Data Handling (GMDH), Genetic Algorithm-Backpropagation Neural Network (GA-BPNN), were applied to model the system and were compared to investigate the best variable combination among 164 variables. It was found that the data need to be first classified in order to enhance the performances. Also, GA-BPNN out performed other methods using only 9 variables. The results were confirmed by engineers and used in FABs to improve the yield by controlling these parameters.