Quality and exposure control in semiconductor manufacturing. Part II: Evaluation

Abstract

The purpose of this article is to test the performance of a heuristic algorithm that computes a quality control plan. The objective of the tests reported in this paper is twofold: (1) to compare the proposed heuristic algorithm (HA) to an optimal allocation (OA) method; and (2) to analyse the behaviour and limitations of the proposed HA on a scale-1 test with a before/after test. The method employed to evaluate this algorithm is based on comparisons: 1. The first test illustrates the method and its sensitivity to internal parameters. It is based on a simplified case study of a product from the semiconductor industry. The product is made up of 1000, 800 and 1200 wafers incorporating three different technologies. The production duration is 1 week, and three tools were involved in this test. The behaviour of the proposed algorithm is checked throughout the evolution of the model parameters: risk exposure limit (RL ) and measurement capacity (P). The quality control plan for each tool and product are analysed and compared to those from a one stage allocation process (named C 0) that does not take into account risk exposure considerations. A comparison is also performed with OA. 2. The second scale-1 test is based on three scenarios of several months of regular semiconductor production. Data were obtained from 23 etching and 12 photolithographical tools. The outputs provided by the HA are used in the sampling scheduler implemented at this plant. The resulting samples are compared against three indicators. The results of these comparisons show that, for small instances, OA is more relevant than the HA method. The HA provides realistic limits that are suitable for daily operations. Even though the HA may provide far from optimal results, it demonstrates major MAR improvement. In terms of the maximum inhibit limit, the HA achieves better performances than C 0, and they are strongly correlated to RL and to the control capacity. The article concludes that the proposed algorithm can be used to plan controls and to guide their scheduling. It can also improve the insurance design for several levels of acceptance of risk.