About WTS Prof. Clair Brown Faculty, Students and Affiliates Research Areas Online Research Reports Working Papers


Second Interim Report
Clair Brown, Editor

12.2 Learning by Doing in Semiconductor Manufacturing

In contrast to many early studies, reductions in direct labor hours are not the source of learning by doing in the semiconductors. Direct labor typically makes only a small portion of variable costs and there is little evidence that direct labor hour requirements fall as the fab gains experience with the process. Instead, it is the elimination of yield losses that results in substantial reductions in manufacturing costs. It is not uncommon for yields for new semiconductor processes to start as low as 10%. In such cases, the cost of scrapped output is initially very high but falls to low levels as yields rise to 90% or higher over time. Since yield improvements are typically based on identifying and eliminating sources of yield loss, the cost reductions are permanents characteristic of learning by doing.

Typically, learning by doing is modeled as though experience itself is sufficient to lower costs. As a result, most studies focus on the choice of an appropriate proxy for experience. Cumulative volume is by far the most common choice, while other variables include time, investment, and changes in equipment. Since yield improvements are the principle source of learning by doing in the semiconductor industry, it is questionable whether experience alone (as proxied by cumulative volume) is sufficient to characterize the learning curve. Yield improvements do not occur as the result of improved labor productivity, but rather as the product of deliberate activities aimed at overcoming sources of yield loss.

The microscopic dimensions associated with semiconductor processes mean that yield problems typically cannot be resolved through repetition. The only case where manufacturing experience is likely to improve yields is line yield losses where wafers are dropped and broken. Here the cause is easily identified, and we would expect that time and experience will improve the wafer handling skills of operators. In contrast, line yield losses where the wafers are misprocessed (incorrect recipe or skipped steps) are not generally observable to the laborers who cause them and have to be identified by engineering analysis. This problem is even more prevalent with die yield losses. Microscopic particles are sufficient to cause failed dice but are unobservable to direct laborers indeed, the operators may unknowingly be the source of the destructive, particles. Parametric problems are similarly difficult to identify, when they occur.

Learning by doing, in this context, is the result of analysis by engineers to discover the sources of yield problems and implement solutions. In the case of parametric processing problems, the role of engineers is most obvious as they work to identify the unknown parameters through experimentation and analysis. Finding sources of microscopic particles is similarly challenging and requires tests and analysis. For line yield losses, engineers work to install equipment and processing systems that reduce the opportunities for mistakes. Even with broken wafers that are easily observed, engineers are often involved in devising and implementing automated wafer handling equipment to reduce mistakes. Thus, as Hatch and Reichelstein (1994) have shown, cumulative production volume and cumulative engineering analysis are the key determinants of learning by doing in semiconductor manufacturing. These variables are not proxies for manufacturing experience, but rather represent the means by which yields and costs improve.

All semiconductor firms perform engineering analysis of production volume to improve yields, but some are much more successful than others. Figure 12.1 shows the defect density trends for several manufacturing processes. It is clear that the levels and rates of improvement in defect densities vary widely between the different processes. The main hypothesis of this study is that some of the differences in rates of learning by doing are a result of differences in human capital and the human resource management practices of the fabs.

While capital equipment influences yields, ultimately learning has to be done by people. The key to learning by doing, and the competitive advantage that accompanies it, is not just acquiring process-specific knowledge, but also acquiring the fab-level knowledge and skills that enhance the ability to learn. The fabs in the CSM study differ substantially in their attitudes about and practices for acquiring human capital. In many cases, equipment operators are expected only to push buttons and load wafers into machines. In other fabs, operators are given responsibilities for equipment maintenance, repair, process control, and limited yield analysis. The fabs invest in the human capital of these operators by training them in math, literacy, statistical process control, and specifics about their equipment and the manufacturing processes. The benefit is not only that operators with more human capital are more efficient and make fewer mistakes, but they become effective problem solvers. This is important for several reasons.

First, including operators in problem solving activities brings more information under scrutiny. Because the operators work most closely with the equipment, they know much about it. They can hear or see problems with the equipment before they lead to serious yield problems. In some fabs, equipment failures are times for operators to take unscheduled breaks, so little is done to prevent these failures or the yield problems that typically accompany them. In other fabs, operators are trained and given responsibility for equipment performance, process control, and yields. They use the information they have from their close association with the equipment to solve problems quickly or to bring in help before the equipment fails. In these fabs, operators often form teams and work together to solve yield problems. With their additional information of the idiosyncratic traits of the equipment, they are an effective resource for preventing and solving yield problems.

Second, investment in operator training is also useful if it frees engineers to work on more difficult challenges. Many engineers complain that too much of their time is spent in "firefighting," i.e., dealing with the myriad of minor crises that temporarily bring production to a standstill, and some of these problems can be dealt with by trained operators. For example, in many fabs, operators are trained in SPC and provided with guidelines for how to identify and solve simple out-of-control situations with their equipment. Similarly, some operators have reported that they spend 30 40% of their time "troubleshooting" yield problems. One fab reported that in the case of a few wafers -with low die yields, the wafers will be given to a group of operators to identify the source and propose solutions. In whatever form, problem solving on the part of operators frees engineers to do more fundamental work such as integrating SPC databases to allow the correlation of SPC data to yields.

The third, and perhaps most important, benefit of investing in human capital is that fabs are able to integrate the entire manufacturing staff into one large problem solving organization. The specific knowledge of more individuals is included in the analysis. The entire organization shares the task and responsibility of yield improvement. The notion of a problem-solving organization is especially important in the context of new manufacturing processes where a fab full of trained employees can tackle the difficulties of learning the new process and equipment with the result of high initial yields that continue to improve.

Operator turnover is the, antagonist of successful investment in human capital and efforts to enhance the rate of learning by doing. When trained and experienced operators leave the fab, they take their knowledge of the processes and equipment with them. Hence, turnover represents knowledge leaving the fab. High turnover creates an additional impediment for learning by doing, because, new employees require extra training and are more prone to make mistakes. The rate of learning by doing will likely fall since the production staff makes more mistakes and requires resources for training that could otherwise be dedicated to yield improvement.

Go to Next Section of this Chapter
Go to Table of Contents for the CSM-HR Interim Report

© 2005 Institute for Research on Labor and Employment. 
2521 Channing Way # 5555 
Berkeley, CA 94720-5555