Skip to Main Content
An approach based on data-mining for identifying invariant objects in semiconductor applications is presented. An invariant object represents a set of parameter (feature) values of a process and the corresponding outcome, e.g., process quality. The key characteristic of an invariant object is that its outcome can be accurately predicted in the changing data environment. One of the most powerful applications of invariant objects involves the generation of robust settings for controllers in a multiparameter process. The prediction accuracy of such robust settings should be invariant in time, features, and data form. The notion of time-invariant objects refers to objects for which the prediction accuracy is not affected by time. Analogous to time-invariance, objects can be invariant in the features and the data form. The former implies that the prediction accuracy of a set of objects is not impacted by the set of features selected from the same data set. The outcomes of data-form invariant objects prevail despite the change in the data form (data transformation). The use of data transformation methods defined in this paper is twofold: first, to identify the invariance of objects and secondly, to enhance prediction accuracy. The concepts presented in this paper are illustrated with a numerical example and two semiconductor case studies.