The ability to rapidly analyze data and provide feedback to improve the production process is a key part of the yield management strategy for setting up a new GaAs fab, writes Paul Burgess of Filtronic Compound Semiconductors Ltd.

Almost two years ago, Filtronic plc purchased a former DRAM fabrication plant in County Durham, UK. The company s aim was to mass-produce GaAs-based chips for mobile telecommunications applications. However, the mass-production techniques developed and refined by mainstream semiconductor manufacturing for silicon have not so far been applied to the processing of what has often been regarded as a niche material. In order to meet market projections and reduce costs, Filtronic needed to successfully implement these mass-production approaches in a new environment, which adds additional complexity to the already challenging process of ramping up a semiconductor fabrication facility to volume production. This feature describes the experiences of Filtronic Compound Semiconductors Ltd in implementing a data-driven yield management system to support the fabrication of GaAs-based RF products. Key business requirements were the rapid delivery of working prototypes to prospective customers, followed by a rapid stabilization of the production process. In spite of the need to utilize manufacturing technology that is out of mainstream semiconductor manufacturing, these requirements were successfully supported through the development and deployment of a custom solution built using SAS technology for data warehousing and exploitation. The need for a data warehouse The telecommunications industry is very competitive, with margins reduced to a bare minimum. The raw materials used throughout the processing sequence are expensive; GaAs wafers are approximately 50 times more expensive than Si wafers, high purity chemicals are used throughout, vast quantities of de-ionized water are required and the manufacturing area has to be maintained in an ultra-clean environment. Furthermore, equipment with a very high capital cost is required to fabricate the wafers. The requirement for the fab was to develop a stable production process, as soon as possible, to enable parts to be sampled to potential customers to support rapid prototyping and help gain market share for Filtronic. The next step was to improve that process, increasing device yields and reducing wafer scrap to predictable levels in order to be in a position to satisfy the demand generated by the prototyping. In both phases the ability to rapidly analyze data from the production process and turn that data into useful information for engineering use was a key part of Filtronic s yield management strategy. The initial requirement for a data warehouse was the storage, analysis and reporting of the electrical test data. The quantity of this data is potentially very large, with the fab capable of producing 30 000 wafers per year, with each wafer containing up to 80 000 individual die, each being tested on the wafer and again when packaged. A key requirement was for the data warehouse to be scalable and therefore "future proof", i.e. capable of handling the quantities of data required to support any realistic business planning scenario. In view of the fact that sources of electrical test data were to be brought on stream one at a time, the data warehouse also had to be modular. Aside from the issues of functionality, covered later, the general solution requirements were below industry average implementation cost, a short implementation cycle, and low ongoing cost of ownership. Fabricating and testing chips The process of fabricating chips on a GaAs wafer is a complex one. It starts with the deposition of epitaxial layers onto the bare GaAs wafer to produce the base layer for device operation. The wafers are then processed in groups of up to 12 (called a lot) through a series of process modules. These deposit and pattern conducting and insulating layers onto the wafer. There may be up to 10 such process modules, so that it takes 10 to 20 days to fabricate the finished device. The final process in the sequence is to dice the wafer into individual die ready for packaging. The wafers used are 150 mm in diameter, and the die size can be as small as 500 500 m2. Therefore, there may be up to 80 000 devices on a single wafer. There are several different process routes, each to produce a different device type for a specific application. Electrical testing forms an integral part of the manufacturing process. It is used both to ensure the operation of the final product and to monitor the fabrication process and ensure stable operation. illustrates the manufacturing and test flow. There are four basic stages of electrical test. The first is the process control monitor (PCM) test, performed at several stages as the wafers are being fabricated in the line. This is typically a test performed at low frequency (DC) on approximately 80 specially fabricated areas on each wafer. The data collected correlates directly with physical parameters of the wafer. For example, the resistance of a metal track relates directly to the thickness of the metal and its patterned width. The second stage of electrical test is performed on finished wafers prior to dicing. This test is performed at RF frequencies, to simulate the operating conditions of the circuits. Again this data correlates with the physical parameters of the wafer and also with the DC PCM test data. After both the DC PCM and on-wafer RF test measurements, the lot is dispositioned and abnormal wafers scrapped. The third test stage is the on-wafer 100% DC testing of the finished die, to ensure that each individual die operates to specification. Each die is electronically tagged as pass or fail. The final stage is testing of the assembled component. The assembly operation consists of placing completed die into a package and is typically done off shore. Testing can be performed in the assembly location or the components can be returned for testing on site. This test is performed at RF frequencies and again ensures that the finished device operates to specification. This test strategy highlights problems as soon after they arise as possible. It also reduces process, test and assembly overhead, by scrapping abnormal wafers as soon as possible. Implementing the data warehouse The project commenced in April 2000 with a period of consultancy from SAS. During this time, the data model and hardware requirements were established, the data warehouse was implemented and priority reporting functions were set up. This phase, completed in October 2000, resulted in a functioning data warehouse and a strong base for further expansion. Since then the application has been continuously developed, with further sources of data added, and reporting functions improved. SAS version 8 was used to implement the data warehouse. A client server solution was adopted to satisfy the requirements for fast data processing and storage of large amounts of data. Filtronic purchased a powerful Compaq DL380 server, with a 2 Gbyte RAM and 85 Gbyte hard drive. This approach provides faster data retrieval rates, along with data management benefits gained from having a central repository of data. Data model A form of a Star Schema Data model was chosen for fast data retrieval and query analysis. This also provides a scalable, extendible long-term solution. DC PCM and 100% DC test data were the first data to populate the database. The data is loaded directly into the database, immediately after test completion, and is split into several tables. These include fact tables containing raw data, lookup tables containing specification limits, test dimension tables containing categorical information (test machine ID etc), summary tables and pass/fail tables. Each table has a header with key fields, allowing the tables to be linked for analysis and reporting. As further sources of data are brought on line, they can simply be added to the model by creating additional tables for each source. The SAS application A front-end application has been written for the data warehouse, using SAS/AF. This allows common tasks to be performed easily and without the need for any knowledge of SAS programming. Tasks are either administration, such as loading the data to the database and deleting unwanted data, or reporting, i.e. the generation of standard reports for engineering analysis or summary reports for management. Electrical tests are performed using automated test equipment, and after measurement the data resides on UNIX boxes that control the equipment. The user initiates data loading by selecting the lot number and measurement stage. Data transfer over the factory LAN then takes place and the data is translated into the various tables. Two reports are immediately generated once the data is loaded. The first is a table summarizing the data for the lot. This is generated in HTML format and is available on the factory Intranet for general viewing. The second is a box plot of test parameter against wafer number, as shown in . This gives a good graphical summary of the data; it shows the level of the data against specification limits and also the spread of the data within those limits. Many of the electrical test parameters are interrelated, so having them all graphed together helps the engineer to identify problems and initiate investigation to find the cause. The combination of these two reports allows a disposition to be made on the lot. The remaining report functions are designed to allow engineering analysis of the data to assist in yield improvement activities. Wafer mapping forms a key part of this procedure. However, with up to 80 000 die on a single wafer, the mapping facility is not trivial. Within SAS version 8 is the ability to generate contour plots (module SAS/GRAPH), which are used to generate wafer maps. (a) shows a map for a 100% DC test, the blue die are pass die and the red ones fail die. The map does not allow individual die to be resolved, but clearly shows the pattern of failed die. The user may zoom in to a section of the wafer, the map will then revert from contour plot to a scatter plot, allowing individual die to be resolved and the probe function to be used to look at data for individual die. (b) shows a DC PCM test map for the same wafer. This is a map of just one test parameter, and because the number of data points is relatively small, a scatter plot covers the whole wafer. The correlation between the two plots in figure 3 is clearly evident, with low values (lighter yellow points) of the DC PCM test parameter corresponding with the pass die area. With this data in front of him, the engineer was able to make a conclusion about the cause of failure and implement a countermeasure. Further reporting functions include SPC charts and capability plots. An SPC chart takes data from multiple device types, sorts it into test order by date, filters out outlying data and plots a normal mean standard deviation chart. This is used as an engineering tool to look at trends in data. Capability reporting is used to look at longer-term trends in data. It is possible to plot histograms of each parameter at each test stage, with capability indices included as a legend, to plot a Pareto chart of capability indices for each test stage, and to make a monthly plot of Cp/Cpk trends for each parameter. An example of the latter chart is shown in . These charts help to highlight parameters with a low Cpk index and to target improvement activities. The improvement in capability indices over time indicates that the process is becoming more stable (the converse is also true!). The benefit of using the SAS data warehouse to generate such reports is the flexibility of data selection to populate the charts and the ease and speed at which they can be created. For more in-depth statistical analysis of the data, SAS/INSIGHT provides a powerful statistical tool for analyzing, plot-ting and reporting the data trends and patterns are easily spotted. One unanticipated benefit of the data warehouse has been the generation of electronic pass/fail maps for the assembly process. Typically in the semiconductor industry, failed dies are indicated by an ink spot. However, the process of inking wafers is costly and messy, it is also not a trivial task when the die size is down to 500 m. The format of the data in the warehouse has made it relatively easy to generate an electronic map to supply to the assembly company. Future plans for the database The short-term goal for the data warehouse is to integrate test data from packaged parts. In the longer term, data from other sources needs to be added. The main source is the factory manufacturing execution system; as wafers are processed through the line, data is entered into this system at each measurement stage. Examples include linewidth measurements at photolithography stages, deposited layer thickness data and epitaxial layer characteristics all physical parameters of the wafers. By adding this data to the warehouse, the existing reporting and analysis tools can be used to look at the data. More importantly, relationships can be established between the physical and electrical parameters of the wafers. This is the next key step in process characterization and yield improvement activities. Conclusions Filtronic made a strategic decision to implement a SAS-based data warehouse early in its total development process as part of its effort to deliver innovative products in the shortest possible time frame. This was considered essential to support a data-driven yield management strategy, seen as pre-requisite for successfully building market share and safeguarding customer commitments. SAS technology allowed us to build a scalable, modular solution that is able to handle the large quantities of data generated by our complex manufacturing process and to convert this data into useful actionable information for our engineers. The benefits of this investment have already been seen with reporting and analysis functions available to engineers at a very early stage in the life of the fab.