Capabilities and lessons from 10 years of APC success
02/01/2004
The current vision for semiconductor manufacturing is that wafer fabs will increasingly use advanced process control — an automated form of tweaking processing on the fly. The concept itself, however, is not new. One branch of its evolution can be traced back 10 years to pioneering work at Texas Instruments. Today, the fab control software from this early program provides great insight into advanced process control capabilities.
Anyone watching the current buzz about advanced process control (APC) in the semiconductor industry may not know that APC was one of the objectives of the Microelectronics Manufacturing Science and Technology (MMST) program begun in 1993, jointly funded by the Advanced Research Projects Agency, the US Air Force, and Texas Instruments (TI). This was a five-year program to develop new semiconductor processes and equipment to produce commercial ICs in less than three days, which was successfully demonstrated with a 1000-wafer run.
Today, the software fruits of this program have been deployed over the intervening ten years in commercial fabs worldwide (see "The evolution of MMST-based APC software"). Indeed, its application provides us with a 10-year view of the capabilities of APC and some valuable lessons about its application.
At the 10-year anniversary of the MMST conclusion, the semiconductor industry now recognizes that there is no "control" in statistical process control (SPC); SPC is just a monitoring, measuring, and reporting activity. APC has become a "new religion" in modern wafer fabs.
Venerable software in action
Today, MMST software technology (in the form of ControlWORKS, ProcessWORKS, TrackWORKS, RecipeWORKS, etc.) is widely used in the manufacture of semiconductor devices, MEMS, and thin-film and GMR disc drive heads. From 1993 to 2003, the aggregate number of licenses for these products now totals 15,300 worldwide, including 2000 licenses for machine control (CW), 1300 licenses for run-to-run process control (PW), 3500 licenses for equipment tracking (TW), 1000 for recipe management (RW), 3500 licenses for fault detection, and 4000 licenses for equipment integration software, the enabler for APC.
It has proved to be a valuable tool for improving equipment Cpk, reducing scrap, eliminating process excursions, reducing manufacturing cost, and increasing fab capacity. This venerable technology includes automated data collection, equipment fault detection, and run-to-run process control — what the industry today calls APC (see figure).
While the MMST form of APC was first installed at TI in 1978, it wasn't until the late 1990s that the worldwide semiconductor industry recognized the value of APC. Early implementations of APC were applied to diffusion operations. Today's popular targets include lithography, etch, and CMP; this is widely accepted because these applications can achieve the largest ROI for an APC implementation.
Lesson on implementation
The concepts of APC are universally understood, but implementation strategies vary by manufacturer. Some device manufacturers prefer internally developed point solutions that differ by process, others prefer an APC implementation that uses a commercially available software application that may be used across all processes in the fab. History is proving that a commercial solution with a large user community is faster to install and requires fewer internal resources.
 A screen shot from ClosedLoop illustrating APC in action, controlling film thickness in a deposition system where deposition rate is degrading. |
We have seen many device manufacturers try to develop point solutions for APC that tend to be just too process-specific and difficult to fan-out across a fab as a common APC solution would. In addition, it is more expensive to develop an internal APC solution than to license a common commercial solution.
Common solutions are easier to integrate with other factory systems and applications. Further, a single APC application reduces training expenses for operators, and process and equipment engineers; and it reduces engineering and IT support costs. With a fab-wide solution, data sharing is easier between all process areas and facilitates easier feed-forward, feed-backward of data to sequential process tools.
Another concept is to acquire the APC capability from an equipment supplier. While this is a convenient way to push the development of APC capability down to the OEM, this concept tends to deliver too many and different APC solutions to each fab that are difficult to implement and manage. What we now expect from equipment suppliers are tools that can provide the equipment, process, and sensor data to support or enable APC.
APC's ROI
Our data shows that successful, fab-wide APC implementations can save more than $1 million/month in scrapped wafer reduction and improve fab productivity by several million dollars/month (see table). In this installation, APC collects data from equipment throughout the fab, interdicts when tools run out-of-specification, and fine-tunes recipes in a feed-forward, feedback fashion that is completely out of the realm of the individual process tools (i.e., APC resides above the tool, not in the tool).
 |
Consider the ROI potential when APC is applied to lithography: The typical goal is to improve Cpk, automate multitool, multiproduct process tuning, and reduce the process engineering effort. The approach is to use APC to calculate overlay parameters based on previous lot measurements and control linewidths, calculate coat spin time based on previous lot measurements, and calculate bake time based on previous lot runs.
We have found that the results for APC on overlay control in a fab with a Canon I4/I5 deep-UV stepper and a KLA 5400 metrology system are in the range of a 75% rework reduction, a 40% Cpk improvement, and a 50% reduction in process engineering effort, ultimately resulting in a 4–6% probe yield improvement. This particular case resulted in $20 million in capital avoidance.
Within the lithography process, APC can automatically control markshift (alignment offset of a layer), scaling, and magnification. In the previous example with the Canon stepper, APC can automatically adjust ~15,000 or more controller parameters across multiple steppers, reticles, and layers.
As another example, when APC is applied to gate-CD etch, the objective is to improve lot-to-lot, within-device, and within-stepper groups Cpk, reduce the processing of nonproduction wafers, and shorten cycle time. The impact of APC on gate etch is consistent. The sources of variation are typically stepper-to-stepper in the photo process and device-to-device variation for etch.
Etch also benefits from APC. Typically, citing data from an actual case study of gate CDs etched on an AMAT 5200, we have seen Cpk improvement from 2.02 to 3.44, the elimination of 8 hrs of cycle time/lot, and significant reduction of test wafer processing. In this example, a single control strategy was used for seven etchers and three target features or geometries resulting in 200,000 parameters (steppers × etchers × reticles) under APC.
When APC is applied to CMP, the goal again is to increase Cpk and to decrease test wafer use, but also to decrease rework. Here, the approach with APC is to use run-to-run control to manage polish time and resulting thickness by creating a process model that uses feed-forward and feedback data from process and metrology systems.
We have found that APC CMP results are dramatic and easy to quantify. For example, in a fab using a BOC slurry delivery system feeding nine AMAT Mirra polishers, we have documented a 25% increase in equipment throughput (which translates to more than $30 million in capital avoidance), a 5–7% reduction in rework, a 40% reduction in scrapped wafers, and a 100–150% improvement in Cpk due to automatically accounting for and dramatically reducing known process variation.
In the manufacture of thin-film heads, APC offers equally impressive results. Consider just two key processes — ion milling used to shape the magnetic poles of the write coil and CMP. Using a model-based tuner to control milling time, fixture angle, and clean-up time, improvements include 40% reduction in sigma, a 15% increase in mean centering, 50% increase in Cpk, and a 5% reduction in scrap.
Using model-based process control to define the read-write separation layer in the CMP process, head manufacturers using our APC have achieved a 55% reduction in sigma, an 85% increase in mean centering, a 100% increase in Cpk, a 10% reduction in rework, and an 8% reduction in scrap.
Lessons learned
While there are many lessons from the applications of APC since its inception, the following 11 should serve anyone involved in the manufacturing of semiconductor devices, MEMS devices, or thin film heads:
- Start simple. Basic linear regression models work well for new APC processes and can provide significant first benefits.
- APC is continuous process improvement. Experience shows that most control strategies are revised many times in the first year.
- APC needs a "buy-in" from all process engineers. They must be able to adapt to a new way of thinking about their processes and understand how they can use these new levers for driving improvements. They must have the ability within their tool set to adjust and improve their models without IT support.
- APC software must be engineer-friendly. Application software that is designed for use by engineers provides an important advantage to a rapid deployment.
- APC can be successfully implemented on process equipment (up to a certain level) without having the benefits of GEM/SECS-enabled software.
- The greatest benefits can be gained with a process using feed-forward and feedback control strategies. This is fundamental to fab-wide, robust implementations of APC that encompass all aspects of equipment communication, data collection, fault detection, run-to-run process control, and fab-wide recipe management.
- Parametric and logistics database systems must be capable of handling real-time data needs. That is, the systems should employ advanced query languages with high volume and fast transactions. APC-enabled fabs will significantly increase the load on these systems.
- APC and application training must be designed to address a wide variety of audience needs and skill levels.
- Additional process benefits can be gained by linking APC-enabled processes and sharing information between control strategies. This eliminates APC point solutions by process as a viable fabwide deployment strategy.
- A thorough analysis of process requirements is key to any APC project. Requirements should include an analysis of current process capability, process verification logic, metrology readiness, parametric database, and tool automation capability.
- APC can be used as a process design tool. This is because APC enables more complex process flows to be handled on a daily basis.
The future of APC
It is clear from our experiences that APC and its components, automated data collection, SPC, fault detection, and model-based run-to-run process control, will become standard manufacturing practice. While the next significant step in device manufacturing will be real-time process control with in situ sensors, that capability is a goal and objective for the semiconductor industry that is still many years from wide deployment in fabs, with significant work to be done in the development of sensor capability. In the meantime, through broad deployment of APC, the industry will make and save billions of dollars per year.
Acknowledgments
ControlWORKS, ProcessWORKS, TrackWORKS, and RecipeWORKS are trademarks of Adventa Control Technologies Inc.
Carl Fiorletta is a director at Adventa Control Technologies Inc., 3001 E. Plano Pkwy., Plano, TX 74074; ph 972/543-1754, fax 972/633-2968, e-mail [email protected].
The evolution of MMST-based APC software
Ten years ago, the Microelectronics Manufacturing Science and Technology Program (MMST) was an experiment into many aspects of advanced IC manufacturing. This included equipment design, single-wafer dry processes, machine and process control software, and a CIM system to manage the overall fab. Until this time, commercial CIM systems were more focused on WIP tracking. The MMST CIM system was more process-centric and included embedded machine controls.
The vision of the MMST program was to define, design and build a semiconductor manufacturing operation with product, technology, and volume "on-demand" flexibility done in an extremely short manufacturing cycle (days, not weeks).
The objectives of the program, particularly in the area of factory and process control, included improved process capability, greater equipment availability and utilization, reduced misprocessing, higher yield, reduced setup and inspection cost and time, less off-line metrology equipment, and faster yield ramp. Ironically, these goals are consistent with the semiconductor industry's present day desire to improve manufacturing productivity and current emphasis on APC.
At the successful conclusion of the MMST program, TI formed a software business unit to commercialize the program's machine, process and fab management software. In 1998 TI sold a majority interest in the software business unit to Gore Technology Group (GTG); GTG formed a stand-alone software business called Adventa Control Technologies to further develop the software products and expand their applications in the semiconductor industry.A screen shot from ClosedLoop illustrating APC in action, controlling film thickness in a deposition system where deposition rate is degrading.