Achieving uniform, systematic APC in multiple fabs

04/01/1999

Customers of semiconductor fabs expect that every chip they buy will display identical performance characteristics. Since the early days of the industry, however, factory-wide systems needed to ensure the process stability and uniformity essential to meeting these customer expectations have been lacking. Both technical and cultural barriers must be overcome by a semiconductor maker at tempting to implement a systematic approach to factory-wide process control.

Over the past few years, we at TI have had some success in achieving the systematic implementation of automatic process control (APC) through out our world wide fabs. Our experiences may provide some in sight to others attempting to deploy similar changes through out their organizations.

Technical barriers were the first to be addressed, as they were of immediate concern to the engineering staff charged with APC implementation. One barrier was a direct result of the nature of IC fabrication processes, which are discrete batch events with batch sizes ranging from one to hundreds of wafers. Each of these "batch" processes must be viewed as both an intermittent continuous-flow process varying over time (run-to-run variation) and as a discrete processing event with variation within the batch (within-run variation). A second barrier was the ever-increasing number of equipment types and processes in a wafer fab, including multiple processes run in a random sequence on one piece of equipment. A third was the lack of sufficient computing power and suitable process control software available on the factory floor.

All of these contributed to less than successful implementation of traditional sampling plan process control based on continuous-flow process assumptions.

There were also cultural barriers to APC implementation, especially in the engineering community. Early staffing decisions generally favored electrical engineers for their device and circuit knowledge. As a result, those engineering groups lacked members familiar with the concepts of manufacturing variation and control - which were strongly emphasized in chemical engineering education but not in the electrical engineering curricula.

Another cultural barrier was a system of rewards and recognition that focused on the individual crisis management skills of process engineers with a knack for correcting process variation. Actually, the engineer who mounted his white horse to slay dragons may have designed the process creating those dragons every time he was not running the equipment! Similarly, wafer fabs were encouraged by the reward system to create local solutions to common problems. There was no recognition for adopting solutions from other fabs.

In the 1980s, however, several strategic decisions were implemented that began to reduce the size of these barriers. One decision involved the recognition by top management that equipment/process modules must be standardized across device families and wafer fabs, and that many of these modules must transfer to at least one higher technology node IC. The expectation that these modules would be re-used by multiple fabs and technology nodes made the time expenditure for process control implementation more advantageous from a fab's cost-benefit analysis point of view.

Another strategic decision allowed the acquisition and continuous upgrade of networked desktop computing resources on the factory floor and at the engineer's desk. This included internally developed reusable APC code based on successful techniques from different fabs. It re moved the need for local code generation, and has now been spun off for external marketing by Adventa Control Technologies Inc. (Dallas, TX).

The upper-level reward and recognition system was also revised so that the composite performance of all wafer fabs combined was measured and included in each fab's review process. This created a strong local interest in removing duplication of effort between fabs and in cross-implementation of the most successful manufacturing techniques (such as APC).

All of these strategies had predictable consequences from a pure technology and financial standpoint, but the impact on engineering and manufacturing personnel was not clearly foreseen. Dealing with human reactions to the implementation of APC has proven to be as important as the creation of APC itself.

Successful process engineers reacted first. Implementation of APC required that they convert qualitative concepts for correcting process variation into quantitative algorithms (or models) through extensive testing. Process engineers faced with this "codification" of their personal insight responded with such statements as "I have too many problems to spend time on development of APC"; "automatic adjustments (no human review before implementation) are too risky for production release"; and "variation should be identified and fixed, not compensated for through some other process adjustment."

All of these statements possess elements of truth, but behind them is the engineers' concern that their expertise and intervention will no longer be the critical factor in keeping the operation running smoothly. Encouragement from direct management is necessary to underscore the personal benefits accruing to the engineer (such as more time to be innovative) after APC implementation.

Once the models were created by the process engineer, production release of the new APC occurred only after local models were validated against global theoretical models and a suitable level of code testing had occurred using dummy data. Even so, during production release, it was not unusual for one or more of the following events to occur:

1. The process Cpk improved even before APC implementation, just from removing sources of variation found by the extensive testing for model construction.
   
2. Some first-order interactions were ignored, regardless of their visibility in the test data, when they did not meet the engineers' prior beliefs about the process.
   
3. Variations occurred in production that confounded the APC software, regardless of the number of prior tests run.

Even the most highly skilled engineers and manufacturing organizations experienced these problems when the APC approach was locally defined from scratch. Importing a successful APC implementation from a sister wafer fab with a similar process was a powerful technique to minimize these problems.

But importation was discouraged by a culture that rewarded local solutions. This manifested itself in such comments as "they run a different technology node, so their APC would not apply"; "the equipment set/wafer diameter/etc. is different, so very little can be learned from their approach"; and "the computer systems are too different for their APC to work in our factory."

These arguments have some merit and must be considered when prioritizing resource allocations. But successful transfers of APC techniques have occurred between very dissimilar fabs.

Shifts in management attitudes are critical for such programs to be successful. First, the reward system must recognize successful implementations as well as successful failures (an implementation that did not work, but created a valuable insight that could be used to avoid future failures). The difference between a failure and a "successful" failure is simple: in the latter, an engineer distributes the details of his attempted implementation, where it failed, and what elements must be changed to assure second pass success. The engineer must have management encouragement to face peer judgment so that the company as a whole will not repeat the same failure again.

Second, upper management must put company-level resources into creating a knowledge database within the company. This database must provide easy knowledge entry, level of confidence indicators in the knowledge contained within each document, and an easy search mechanism to find documents matching a problem. Management must then reprioritize the engineering staffs' tasks so that they have the time to insert knowledge gained in the database (delaying initialization of the engineer's next project). Without such a database, it becomes very difficult to turn an APC implementation failure into a "successful" failure.

Third, local management must be willing to give engineers time to share interesting successes and failures. In addition to database time, they must have time and travel support to attend annual APC internal conferences, key equipment user group meetings, and even temporary assignments to other fabs. Opportunities must be created for unstructured discussions between personnel from different fabs with common problems.

Fourth, all management must formally recognize and reward both the sending and receiving organizations when a solution to a common problem is successfully imported from one fab to another. Transfer of process knowledge is difficult, requiring resources from the sending fab (with no improvement in their own operations) and the removal of a "dragon-killing" opportunity for the receiving fab's staff.

Since Texas Instruments has implemented these management techniques, improvements in yield, total cost of ownership, and Cpk have been achieved as barriers to APC have been overcome. These techniques have also enabled rapid cross-company transfer of other beneficial manufacturing practices. All of the techniques must be implemented if systematic APC is a company's goal.

Author

Kenneth G. Vickers is on educational leave of absence from the Sherman Wafer
Fab at TI, where he is engineering man ager. He currently serves as director of the Arkansas Center for Electronic-photonic Materials Innovation, 3179 Bell Engineering Center, University of Arkansas, Fayette-ville, AR 72701; ph 501/575-8412,
e-mail [email protected].