This Small Business Innovation Research (SBIR) Phase II project will create a first Real-time Economic Sampling (RES) system for quality intensive high-tech manufacturing. Focusing initially on semiconductor manufacturing, the system will allow defect inspection sampling to be optimized and adjusted many times per day across all process tools, steps, and products. This minimizes, among other things, a manufacturer's economic risk of producing bad products by adaptively focusing sampling where it gives the greatest return. The complexity of the underlying interdisciplinary models have hindered the realization of RES. Recent advancements by the applicants showed feasibility that this research will build on. There are three main research categories, 1) Risk & Optimization, 2) Cycle-Time, and 3) Yield Modeling. Categories 1 & 2 build on Phase I feasibility results where fast, yet accurate, approximations of complex factory models were established. Risk & Optimization modeling will introduce near-term risk due to factory floor events into traditional (long-run average) risk models. Cycle-time research will refine the approximations for queues with heterogeneous servers. Yield modeling will introduce novel real-time yield approximations needed for RES. The included implementation tasks will then create the first factory-ready RES system that captures the necessary intricacies of semiconductor manufacturing.
The proposed RES system is a "Green Manufacturing" enabler that creates a new market as it will be the first to recoup inefficiencies present in many process control operations. The initial focus is on semiconductor manufacturers who are limited by time-consuming off-line analysis, rigid sampling rules, and risky ad-hoc sampling adjustments. Meanwhile valuable products are being wasted due to out-of-control incidents and/or under-utilized inspection capacity. The proposed system can change this. An RES system belongs to a fast growing segment of the semiconductor industry: process control. Other markets can also be pursued and a societal impact due to reduced waste can become widespread.
Introduction This Small Business Innovation Research (SBIR) Phase II project created the first Real-time Economic Sampling (RES) system for quality intensive high-tech manufacturing. Focused initially on semiconductor manufacturing, the system allows defect inspection sampling to be optimized and adjusted many times per day across all process tools, steps, and products. This minimizes, among other things, a manufacturer’s economic risk of producing bad products since it is now possible to adaptively focus quality control sampling to where it achieves the greatest return. In summary, the goal of the RES system invented is to essentially enable "predictive quality control", i.e. enable scarce quality control resources to dynamically adjust and "look" for manufacturing problems in areas they are more likely to take place (this is similar to "predictive policing" where the police plans police car patrols based on both historical and recent crime reports). In semiconductor manufacturing, processing tools are said to undergo "excursions" when something goes awry and they start to produce at a lower yield. Sampling is necessary to detect these excursions since the inspection equipment is very expensive making it impossible to have enough capacity to inspect all units exiting process tools. The RES system invented helps semiconductor manufacturers to continuously adapt the sampling towards higher risk tools, operations, and products. Inventions made Before this invention, the complexity of the underlying interdisciplinary models had hindered the realization of RES. To address the limitations, innovations were made in three research categories: 1) yield risk estimation & sampling optimization, 2) cycle-time & queue modeling 3) yield modeling Innovations in categories 1 & 2 above built on Phase I feasibility results where fast, yet accurate, approximations of complex factory models were established. The additional innovations made in Phase II allow near-real-time sampling optimization to be achieved (previously took 24+ hours) for the shorter horizon objectives of an RES system. Modeling innovations for yield risk estimation introduced near-term risk into traditional "long-run average" risk models. The result is an ability to estimate the excursion risk a process tool has at any given moment, as opposed to the long-run average excursion risk for, say, a three to six month period. Innovations for sampling optimization took place in parallel to all other R&D described here. The end result is an algorithm that can optimize quality control sampling rates as a function of all the other components discussed here, i.e. yield risk/yield models and cycle-time/queue estimates. The cycle-time innovations made provide queue length approximations for queues with "heterogeneous servers". This is important since the queue time in front of quality control equipment contributes greatly to excursion detection times (being able to model "heterogeneous servers" means more accurate modeling of factories since it is common to have a variety of non-identical tools inspecting the same materials). Yield modeling innovations made enable excursion risk on process tools to be represented as yield estimates needed by the sampling optimization. In addition to the research itself, all of the above innovations were implemented into an enterprise software system that can be installed in a semiconductor factory. The implementation work addressed various additional components, e.g. factory data feed processing, estimation of various statistics in real-time, and the creation of an intuitive GUI/dashboard. Broader impacts of the innovations The RES system created under this grant is a "Green Manufacturing" enabler that creates a new market since it will be the first system to recoup inefficiencies present in many process control operations. The initial focus is on semiconductor manufacturers who are limited by time-consuming off-line analysis, rigid sampling rules, and risky ad-hoc sampling adjustments. Meanwhile valuable products are being wasted due to excursions and/or under-utilized inspection capacity. The system invented changes this. A report from collaboration with SEMATECH/ISMI showed a potential benefit of $3 to $30 million/year to a factory that uses RES for defect inspection. Scientifically, the R&D conducted broke new ground in the field of economic statistical process control while also introducing new fast approximations useful in queueing theory and general factory modeling. Marketwise, the RES system belongs to the fastest growing segment of the semiconductor industry. Spending on process control went from 10% in 2000 to 19% in 2007 (Source: Dataquest). Other markets can also be pursued and a societal impact due to reduced waste can become widespread. Since the marginal benefit of efficiency improvements is higher for factories in higher cost regions, the software invented plays a role in keeping U.S. high-tech factories globally competitive.
