An Artificial Intelligence-Based Framework for Burn-in Reduction in the Semiconductor Manufacturing Industry

In the semiconductor manufacturing industry, burn-in is performed to screen out latent defects, which can cause failures during early-life stages. However, burn-in is expensive and time-consuming, since it requires extensive testing under accelerated stress conditions, such as high temperature. As early-life failures are originated during the different stages of the manufacturing process, we develop a framework to estimate the quality of a production lot by exploiting different sources of data collected from different production machines. Specifically, we consider (i) signals measured from the machines used in semiconductor production; (ii) wafer map images collected by performing probe tests on the dies of the processed wafers; and (iii) results of electrical tests performed prior to burn-in. With regard to the exploitation of the data in (i), a method for detecting anomalies of the most critical production machines based on long short-term memory (LSTM)-based convolutional neural network (CNN) has been developed. With regard to the exploitation of the data in (ii) and (iii), two methods for predicting the production quality have been developed. The former is based on the combination of principal component analysis (PCA) and one-class support vector machine (OCSVM), whereas the latter is based on support vector regression (SVR). The results of the applications of the developed methods to real production data show that the quality of the production can be effectively estimated and, therefore, used for a data-informed decision on the number and type of burn-in tests to be performed.