Yield Modeling with Rule Ensembles

In this paper we introduce the application of a new statistical modeling algorithm called rule ensembles to the problem of yield-loss characterization. Yield loss modeling is viewed as a regression or classification problem, and a model is constructed as a linear combination of simple rules derived from the data. These rule ensembles have been shown to produce predictive models competitive with the best methods. In addition to their high accuracy, however, these rules are easy to understand. Similarly, the degree of relevance of each rule, and its respective variables, can be assessed. The algorithm also provides methodology for automatically identifying those variables that are involved in interactions with other variables, and the strength and degrees of those interactions. To illustrate the interpretation advantages of the method, an analysis on semiconductor manufacturing data is provided.