Advanced Process/Equipment Control

Eliminating scrap through process and equipment control based on interactive learning

Joyce Hyde and Pam Ward, Ibex Process Technology, a division of NeuMath

The semiconductor industry has increasingly recognized the importance of advanced process and equipment control to maintain manufacturing integrity and prevent yield loss. This article examines a process excursion that affected a group of etch tools at National Semiconductor's facility in South Portland, ME. While initial indications pointed to a less-than-optimum recipe, closer analysis revealed that the tools suffered from a chronic temperature problem.

To correct the faulty process variable, the fab installed the Dynamic Neural Controller from Ibex Process Technology, a division of NeuMath (Haverhill, MA). Employing continuous interactive learning, the controller eliminates the difficulties typically encountered with standard fault-detection systems. It can recommend both recipe adjustments and scrap-preventing maintenance actions. At National, the system offered several alternatives to solve the temperature excursion, emphasizing those with the lowest cost or risk. Implementing the solutions resulted in improved tool performance and reduced maintenance interventions. The installation of the controller resulted in a 60–75% reduction in process aborts and the elimination of major scrap events.

Intelligent Process Controller

Traditional control methods are limited in their ability to respond to complex real-world scenarios.¹ In place of such methods, biological techniques in which intelligence emerges as a program learns from its experiences are gaining momentum in the semiconductor industry. For example, etch tools have long been the "black box" of semiconductor processing because the etch process is complex and generates huge amounts of information. By applying biological techniques such as neural networks and genetic algorithms in a process controller, patterns showing both failures and solutions can be visualized easily, enabling process engineers to respond accordingly.

The controller installed at National Semiconductor uses such biological techniques. It models the health of a tool and predicts process results using trace data, metrology data, and maintenance logs. The controller offers several design features:

• First, it is adaptive, requiring minimal human intervention. After it has been initialized by learning from the data it has processed, the controller continually updates itself, maintaining accuracy so that the model does not become obsolete along with old process techniques. This feature is critical for modern fabs that constantly update their processes.

• Second, because it is model based, the controller can make near-real-time predictions and recommendations. It can identify process deviations and trigger alarms in seconds rather than days, as required by traditional process controllers.

• Third, the controller identifies several courses of action to improve results. It weighs them according to their cost or risk to wafer quality and the cost or risk in performing them. For example, maintenance actions that require opening the chamber are costly and risky because of the time required to perform them and the danger of leaks. The controller provides process and equipment engineers with a list of actions available to them on a regular basis, enabling them to determine whether a procedure is beneficial enough to be performed at a given time and to choose the one that fits with the current state of production. For example, if a tool has only one lot to run before the end of a shift, the controller can instruct the engineer to run a conditioning wafer instead of performing a chamber clean to reduce the risk of a process excursion, after which the tool can be stopped and cleaned.

Accurate Model Performance

Model accuracy is essential to a model-based process or equipment controller. Inaccurate models not only fail to control, but also introduce additional process variability—which can be fatal. The controller constructs individual neural networks for each quality parameter (quality metric) of importance to the process. It calculates a normalized accuracy for each network to compare different metrics. For each quality metric, the possible range (minimum to maximum value of the metric) is divided into eight bins based on the process-limit setting. Accuracy is defined as the percentage of times that the observed results fall into same bin as the predicted results.

The rationale behind this definition of accuracy is that if a quality parameter moves from within to outside the safety limits, the process suddenly becomes very costly, because parameters outside the safety limits usually trigger alarms or even tool shutdowns. Therefore, parameter values within a safety limit should be treated differently from those outside a limit, although there may be only a very slight difference between the two. Based on accurate prediction derived from the bins, process engineers can determine whether a process is within the control limits, although a predicted value may not match the observed measurement exactly. Experience using this method has shown that for small sample sizes, 35% model accuracy provides a fairly good match between predicted and observed results. For large sample sizes, accuracy can approach 100%.

Figure 1 shows the predicted model values versus the observed values for mean final inspection critical dimension (FICD) in two real fab production recipes. Although the model accuracy was 48%, the rms error was only 0.025. The total sample size was 628 points. Visual inspection indicated that the predicted mean FICD matched the observed values well. Although the two recipes alternated during the production period, model performance did not degrade when the tool switched from one recipe to the other. On the five etch tools modeled, model accuracy ranged from 33 to 95%, with an average accuracy of 63% and a standard deviation of 24%. The success of subsequent process recommendations is rooted in model accuracy.

Figure 1: FICD mean fitting for recipes 18 and 46. The pink and blue boxes are the predicted and measured values for recipe 18, respectively, while the red and black boxes are the predicted and measured values for recipe 46, respectively.

Process Excursion

At National, one etch chamber required a recipe parameter correction. As illustrated in Figure 2, the controller reported that the bottom electrode temperature for recipe 57 needed attention (the data plots above the threshold line indicate a condition of high urgency). The concept of urgency denotes that certain actions must be performed to keep the process in control. Urgency is based on a cumulative sum (CUSUM) statistic and is applied to both maintenance and process parameter adjustments.²

Figure 2: Urgency calls for an adjustment to the bottom electrode temperature for recipe 57. The green line is the urgency value. A prolonged period of urgency (more than 25 wafers) above the threshold (the red line) indicates that the temperature needs immediate attention.

As bottom electrode temperature was plotted over time, it became clear that temperature was shifting slowly toward the acceptable limit (see Figure 3). Although the parameter was still within limits, the deviation from the target (45°–55°C) affected wafer quality. This effect was captured by the control system's multivariate neural networks. Traditional statistical process controllers would not have captured the effect and would have considered the process to be in control. The controller at National, however, not only identified the problem and triggered an alarm, but also offered recommendations for bringing the etch process back under control.

Figure 3: Plot of bottom electrode temperature over time for a high-volume recipe shows an increasing temperature trend. The upper and lower temperature limits are 55° and 45°C, respectively.

Sometimes the controller recommends that multiple remedies be implemented at the same time because making a single process change may move one or more parameters closer to the target while moving others further away. The optimum solution is to move as many parameters as close to the target as possible at the lowest cost and with the lowest risk to the wafer. Although the best choice may be obvious, process engineers often find that the combined remedies recommended by the controller are helpful when the situation has never been encountered before.

In this case, the controller recommended two actions to fix the problem: filling the thermal control unit (TCU) and reducing the bottom electrode temperature. While reducing the electrode temperature alone would have solved the temperature problem, as Figure 4 shows, performing an inexpensive "fill TCU" action further reduced the defect density and brought the chamber endpoint closer to its target.

Although the temperature drift associated with the bottom electrode became an urgent problem only for recipe 57, this was not simply the case of a less-than-optimized recipe. Typically, a high-urgency condition for all recipes in a family indicates that recipes are less than optimized. However, a high-urgency condition across all recipe families may indicate tool-related issues, since it is highly unlikely that all recipes will run in a less-than-optimized state. A temperature trend for a single recipe that results in poor wafer quality may indicate that a maintenance action will correct the problem, although the controller has not yet "learned" this.

Real-Time Results

Five etch tools at National have been running with the controller since February 2003. The installation of the controller has led to a dramatic improvement in tool maintenance operations. Results from the installation demonstrate that significant savings are possible by performing the right maintenance actions at the right time.

To demonstrate the benefits of the controller, maintenance actions performed on the tools two months before and after the installation were compared. For all five tools, aborts had decreased by 56% at the end of the test period.³ A detailed comparison is listed in Table I. Of the aborts that occurred after the installation, 55% were preceded by high-urgency maintenance-action warnings from the controller, as shown in Table II. Apparently no specific maintenance actions were performed, regardless of the warnings. If preventive actions had been performed, more aborts might have been eliminated.

This hypothesis is proven by the fab's production data. Aborts have decreased as the fab has increasingly followed the controller's recommendations. For example, one of the etch tools experienced 11 aborts in the first six months after the installation of the controller. During the next six months, however, only one abort occurred. That improvement has continued: the tool has experienced only two aborts during the most recent 11-month period, as indicated in Table III. Most importantly, major scrap events (defined as scrap on continuous multiple lots) have been eliminated since the controller was installed.

Although the controller recommends maintenance actions to improve overall process results, it does not recommend excessive maintenance. The two tools that followed the controller's recommendations most closely experienced 5.5 fewer maintenance interventions per tool per month—an 11.4% reduction in maintenance interventions, or the equivalent of 33 fewer interventions per tool per year. The interventions that did occur shifted from high-cost/high-risk activities to low-cost/low-risk ones. These observations demonstrate that the controller can predict when maintenance actions and process changes are required and can determine when unnecessary maintenance actions are being taken.

The controller not only reduces tool maintenance costs but also improves process capability. There are two reasons for this. First, by ensuring that the tools function optimally, the controller enables processes to run more smoothly and tightly. Second, because there is no theoretical or practical separation between process and maintenance parameters, the controller's recommendations are based on a comprehensive evaluation of both maintenance actions and process inputs. Therefore, the controller recommends adjustments in process parameters from time to time to offset the excessive impact of maintenance actions. Process parameter adjustments can aid new recipe development, since the controller can serve as a simulator of the process tool. It can also be used to fine-tune recipes on aging tools, which is of increasing interest to fabs. Much critical information about the state of the process and the tool can be gleaned from the controller's process parameter adjustment recommendations.

Conclusion

The net impact of the controller on tools at National Semiconductor has been significant. The amount and cost of maintenance interventions have been reduced while their effectiveness has increased. Scrap costs have been all but eliminated. Process capability has become tighter and throughput higher. In effect, National's tools meet more-stringent specifications while offering higher throughput and a lower cost of ownership. In other words, overall equipment effectiveness has improved.

In addition to providing these benefits, the controller can model and monitor many more variables than were previously considered by engineers at National. This advantage has allowed them to uncover complex interactions that previously have gone undetected. Understanding these interactions can provide significant insights into which variables affect process and tool health. Interpreting these interactions requires in-depth knowledge of processes and tools. Efforts are under way to interpret these complex fab interactions.

The controller has been installed on tools outside the etch area, and tests are being conducted to evaluate its effectiveness. The expansion of the controller's capabilities is also under investigation, including in the areas of sensor data extraction and multiple process control.

Acknowledgments

The authors would like to acknowledge Jill Card and An Cao from NeuMath and Paul Fearon, Brett Getsinger, Kipton Hayes, Ken Swan, Scott Hopkins, and David Tucker from National Semiconductor for contributing to this article.

References

1. G Mone, "Could Robots Take Over the World?" Popular Science 265, no. 2 (2004): 59.

2. J Card et al., "Beyond AEC and APC—Wafer Quality Control," unpublished manuscript.

3. J Hyde et al., "The Use of Unified APC/FD in the Control of a Metal Etch Area," in Proceedings of the 15th Annual IEEE/SEMI Advanced Semiconductor Manufacturing Conference and Workshop (Piscataway, NJ: IEEE, 2004), 237–240.

Joyce Hyde, PhD, is director of applications engineering at Ibex Process Technology, a division of NeuMath (Haverhill, MA). With more than 10 years of experience in semiconductor fabrication, she has worked for Quantum Corp. and Digital Equipment Corp. Hyde has coauthored 15 scientific papers. She received a PhD in materials engineering and electrical engineering from Rensselaer Polytechnic Institute in Troy, NY. (Hyde can be reached at 978/556-0367, ext. 116, or jhyde@neumath.com.)

Pamela Ward is chief operating officer at Ibex. With more than 20 years of experience in the field of plasma physics, she joined the company's management team in July 2004. Before joining the company, Ward was cofounder and vice president of R&D at Peak Sensor Systems. She has also held various senior positions at Sandia National Labs in Albuquerque, where she developed sensor technology aimed at process control applications in the semiconductor industry. She holds several patents and received a TIE degree in materials science from Sandia National Labs. (Ward can be reached at 978/556-0367, ext. 121, or pward@neumath.com.)