Process capability indices are widely used to check quality standards both at the production level and for business activity. They consider the location and the deviation from specification limits and targets. The literature contains many contributions on this topic both in the univariate and the multivariate context. Motivated by a real semiconductor case study, we discuss the role of rational subgroups and the challenge they present in the computation of capability indices, especially when data refer to lots of products. In addition, our context involves a mix of problems: unilateral specification limit, nonsymmetric distribution of the data, evidence of data from a mixture of distributions, and the need to filter one component of the mixture. After solving the previous issues and because of the peculiar characteristics of semiconductor processes based on the so called "wafers," we contribute to the literature a proposal on how to compute capability indices in the case of heteroscedastic spatial processes. With a generalized additive model, we show that it is possible to estimate a capability surface that allows the identification of regions expected to not be fully compliant with the desired quality standards. K E Y W O R D S dry-etching semiconductor processes, mixture distributions, process capability indices, sampling network 1 | INTRODUCTION Process capability indices (PCIs) are well-known tools to estimate the mean-deviance performance of key product characteristics with respect to both targets and specification limits. The main aim of PCIs was originally to evaluate the instant quality of a process given specification limits and prior to manufacturing. Next, because of their simple interpretation, they have also become a fundamental tool for commercial activity. Nowadays, PCIs are unavoidable instruments for managers and engineers; hence, when assumptions upon which these indices are grounded can be violated, much effort has been made to extend and adapt them to such situations (see Montgomery 1). A flow of contributions is available in the literature for both univariate and multivariate cases. The capability indices that researchers have proposed are so many that, since the masterful work of Kotz and Johnson, 2 the word "avalanche" has been coined. A detailed and somewhat critical review on PCIs is available in Kotz and Lovelace, 3 whereas Vännman 4 proposes a unified approach. A thorough review of contributions is reported in de-Felipe and Benedito 5 and Bong-Jin and Kwan-Woo 6 among others). Most of the papers focus on the assumption of normality of the random components, but how capability indices could be efficiently estimated when non normal distributions are involved remains
The monitoring of spatial production processes typically involves sampling network to gather information about the status of the process. Sampling costs are often not marginal, and once the process has been accurately calibrated, it might be appropriate to reduce the dimension of the sampling grid. This aim is often achieved through the allocation of a brand new network of less dimension. In some cases that is not possible and it might be necessary the selection of a subgrid extracted from the original network. Motivated by a real semiconductor problem, we propose a method to extract a monitoring subgrid from a given one, based upon grid representativeness, accuracy, and spatial coverage of the subgrid and, if available, by expert knowledge of the weights to be assigned to those areas where production may need greater precision. Discussion is mainly focused on circular spatial domain, since, in microelectronics, the basic production support, called wafer, is a circle. Straightforward generalizations to different spatial domains are possible. Furthermore, conditionally upon the availability of experimental data, we check the loss of accuracy by fitting a dual mean‐variance response surface on the reduced grid. Joining the latter information and the criteria used to select the subgrid, we provide additional guidelines on how to fine‐tune the subgrid selection. Real case studies are used to show the effectiveness of the proposal.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.