Semiconductor chip industry has more than 50 years, as more and more widely applied, the human dependence of electronic products also more and more deep, the semiconductor chip industry has become increasingly important role. Intel (Intel co-founder Gordon Moore (Gordon Moore) in 1965 Moore's Law (Moore "s Law), that the progress of process technology, can every 12 months in the same units in the area of wafer (wafer) into the double the number of transistors (transistors). The development so far, constantly miniature semiconductor components, line width has entered into 16 nm, a size of a fingernail integrated circuit (IC) can be put in more than billions of transistors, the line more than one over ten of a human hair.
Integrated circuit manufacturing is to design a good circuit, through repeated exposure, development, such as ion implantation, etching hundreds of complicated manufacturing process, take as many as 30 layer above, each layer of the circuit are accurate forming in the round cake pieces such as chips on the wafer, and finally with the back of the packaging test for each chip (chips), and the production cycle time more than a month.
Chip manufacturing depends on the tip of the advanced lithography technology, effectively managed talent and precision equipment is very expensive, and the chip industry supply chain is dependent on its background powerful logistics safeguard system and upstream and downstream of supply chain collaboration.
Into the era of nanometer process, semiconductor chip process is more long, complicated influence variables is higher and higher, more technical threshold, research and development costs and capital spending capacity to form a double burden, taken in the process of production or abnormal, could cause the qualification rate of loss or even discarded.
At the same time, along with the diversification of IC products with short life cycle is getting more and more, how to promote quickly with the analysis of large data, nanometer process qualification rate, and through the formation of upstream and downstream of the effective system of logistics supply chain, has become the international competition strategy of semiconductor companies.
Hsinchu, tsinghua university (1996) established the decision analysis laboratory, the application of data mining and decision analysis method to study how to improve the qualified rate of semiconductor, and in view of the low qualification rate of wafer sort, then mining caused low qualified rate of the process, product category, equipment, time, possible reasons, such as combination algorithm, information technology and the graphical user interface, the hoisting system of the "qualified rate".
Study we also found that when people are thinking about how to improve the percent of pass, typically focus on solve the problem of abnormal process and equipment, but the essence of "qualified rate" should be on a wafer output can be sold most of grain. As a result, we established the comprehensive benefit of Wafer (Overall Wafer Effectiveness, OWE) index architecture, and put forward the use of data analysis, to change the grain arrangement Wafer qualified rate increased innovation idea.
We sort out the data mining was used to optimize design of IC wafer output size guide (gross die advisor), no matter how much experience can quickly make engineers decided that the best way of configuration, wafer exposure and effective certificate can increase grain output, improve work efficiency and the equipment efficiency, and reduce customer complaints, average benefit estimates that up to nt $425 million each year, the technology has been imported 8 inch and 12 inch TSMC factory, in order to service its downstream customers.
Since 2003, the author will be the start of the TSMC complex practical problems in a mathematical model, establish can always empty environment transformation mode of decision analysis, and import the data mining to reduce the production cycle time (cycle time) to enhance the productivity of the method.
Into the era of consumer electronics, semiconductor chip product value as time rapid depreciation, so the time to market, and a shortened the production cycle time is very important. In addition, as a result of the semiconductor production pattern is very complex, so the traditional production management theory can only handle small workstation.
We use semiconductor manufacturing huge amounts of data, analyzed the wip level and online waiting time and impact factor, each workstation in products online to find out the ideal relationship between water level and output, through the macroeconomic regulation and control mechanism to maintain the balance of production system and machining process smoothly, effectively shorten the production cycle.
TSMC has the fab automation development, divided into anthropomorphic, unmanned, superman three stages. Learn first is to use computer and equipment, the second is the mechanical automation to replace jobs, the last is to develop a rally all intelligent manufacturing system. Not only can let the system automation, but also "smart" to know how to judge and decision-making, beyond the ability of ordinary people. This is not only the future trend, is also a great challenge.
Semiconductor nano process variation to grow, the technical difficulty and fully automated 12 inches fabs capacity more than one hundred thousand pieces, online at the same time in a dozen process recipe parameters (recipe) the production of various products, every piece of wafer after hundreds to thousands of repeated cycle of manufacturing process, each workstation has several to dozens of precision of the reaction chamber (chamber) can choose, the process of production can be read as time tens of thousands of kinds of real-time monitoring data, nearly all online sampling measurement value (metrology), as well as a few hundred in different locations on a wafer measuring electrical test parameters, combined with integrated circuit complex mode of production, make the data in addition to the common characteristics of 4 v big data, also is a large number of (volume) and diversity (variety), rapid change (velocity) as well as the authenticity, veracity and so on, and the main effect is not obvious, data distribution is not balanced, before and after the process such as the interaction of the complex challenges.
With continuous miniature semiconductor processes, on the other hand, challenges the physical limit, permissible error are also shrinking, make it hard for even a senior engineer, to professional knowledge and experience, alone or traditional statistical analysis method, from the big data quickly find out the cause of the abnormal process.
Although commercial statistical software to support large data analysis gradually, but due to a lack of demand for semiconductor industry and characteristics of application modules, has affected the general engineer 'willingness to use. Therefore, beginning in 2011, TSMC promote existing Engineering Data Analysis (Engineering Data Analysis) system upgrades, and began to "intelligence to assist in advanced nano manufacturing process increase percent of pass" university-industry cooperation plan.
Semiconductor chip according to the analysis of the big difficulty is that the semiconductor manufacturing data generated in each stage has a close relationship. Therefore, must consider the timeliness, clustering, data linkage, rather than abuse of computer computing power to make data fishing (data dredging).
TSMC currently working with tsinghua university, combines data preparation techniques and analysis technology, through data analysis technology automatically accumulating large data into valuable information, combined with the experience of decision makers and ability, become exclusive manufacturing enterprise intelligence, and join the interdisciplinary field of biological information talent, to join TSMC rtvu data analysis of the relevant department.
Laboratory through the cooperation with TSMC development units and experts in the field of close cooperation, combined with the theory and domain knowledge as a comprehensive data analysis, to build understanding and mastering the complicated semiconductor manufacturing system, the integration of large data analysis, data mining, decision analysis and graphical technology, method and information system, the development of data mining architecture and algorithm is suitable for semiconductor data characteristics, finally successfully established different analysis techniques such as multivariate accident analysis and diagnosis module, so as to shorten the learning curve of user, support engineers for subsequent data analysis and other professional judgment, substantially improve the quality of decision making of engineers, accelerate the qualified rate of ascension.
