Process Equipment—Furnaces

Using automatic fault detection to improve diffusion furnace performance and
reduce wafer scrap

Mark Yelverton, Brian Cusson, and Tom Timmons, AMD; and Kevin Stoddard, SEMY Engineering

Increased demands on semiconductor manufacturing are driving the development and implementation of advanced control strategies. Partnerships between equipment suppliers and semiconductor manufacturers make these improvements possible through the development of open hardware platforms and data management solutions. Adding sensors to tools where needed and generating sensor data are critical, regardless of whether model-based, run-to-run, or automatic fault detection (AFD) control strategies are implemented. Accurate, reliable in situ information is the foundation on which all control systems should be built.¹

Ideally, real-time, wafer-level in situ data provide the most direct process control loop, as long as the data are reliable and cost-effective. However, wafer-state and process-state sensors are often not available for many tools, so that control information must be based on equipment-state sensors and delayed measurements. Figure 1 shows a complete advanced control diagram detailing such control loops.^2,3 These indirect feedback loops can increase cycle time and are more labor intensive than wafer-state and process-state sensors. Additional equipment-state sensors are often added to a system to detect process parameter variations, increasing the difficulty of data management. Drift and noise at the equipment level are also major concerns, and if not detected and corrected quickly they can affect yield, increasing manufacturing costs.³ Control depends on information, and data that are unreliable or difficult to use become noise.

Figure 1: Complete advanced process control diagram showing control loops.

This article describes the AFD method of in situ data management used to monitor equipment-state sensors on diffusion furnaces at AMD's Fab 15. This control system is useful for monitoring and reducing sensor drift and noise, enabling corrective action to be taken before process- and wafer-state variables are affected. Combining the AFD system with a concise approach to system design for maximizing data integrity has led to improved furnace performance and reduced wafer scrap, providing a return on investment in less than a year. The AFD system discussed here converts system data into information and handles sensor drift. In situ equipment-state data measurements are reduced through programmable algorithms to a go/no-go type of run-to-run control. Data from any input, at any time within a run, can be converted graphically to a statistically tracked data point with alarm limits that alert production to a system change before the process or wafer state is affected in the next run.

Design and Implementation

A partnership was formed in 1995 between AMD (Sunnyvale, CA) and SEMY Engineering (Phoenix) for the
development of advanced control solutions. The first step in this partnership was to maximize the data integrity of the furnaces by upgrading the furnace sensors and control hardware. To handle the increase in information, the SEMY furnace control and supervisory system was selected for the furnaces. The vendor designed a new hardware I/O interface that is open and flexible, which allows for expanded digital and analog signals. With this solid foundation in place, the furnaces were equipped with AFD to further strengthen data integrity. Along with AFD, data integrity was maximized through sensor choice and system design. Methods for improving the data from various sensors on hot-wall diffusion furnaces were implemented. On this platform, model-based and run-to-run control were later developed. This section focuses on the sensor modifications and the use of sensor-generated data with AFD.

System Design for Maximizing Data Integrity. The best way to improve a control system's performance is to first reduce variation and noise wherever possible. The proper selection, use, and maintenance of the system's sensors are critical first steps toward ensuring that the correct information will be present in the data. In hot-wall furnaces the main processes are CVD, oxidation, and diffusion/anneal. For CVD and oxidation the main results are deposition or oxidation thickness and uniformity. For diffusion and anneal the thermal effects of dopant depth and film composition are critical. Few if any process-state or wafer-state sensors are readily available for these processes, requiring that equipment-state sensors be used to control the main variables of gas flow, pressure, temperature, and time.

When determining how to measure and control these variables, several factors must be considered, such as intrusiveness, accuracy, sensitivity, repeatability, cost, maintainability, sample rate, and failure rate. These factors are affected by the types of sensors selected as well as by the range of the sensors, their placement in the system, their ability to be calibrated in situ, and system design. Many of these factors have trade-offs, and the weight given them is critical and different for each variable to be measured. In all cases the ability to perform in situ calibration checks wherever possible is particularly crucial. Sensor removal for checking calibration off-line can damage the system and the sensors.

Pressure. In CVD processes pressure is a critical variable that affects the deposition rate and film properties. In the system under investigation a capacitance manometer in conjunction with a throttle valve is used to control pressure. For improved accuracy, two manometers are used, the first of which is a large-scale 0—1000-torr unit for gross pressure measurement and the second a 0—1-torr unit for more accurate measurement around the control point. Both units are heated and placed at the front of the system, away from the main flow of gas, to reduce unwanted deposition on the sensors and to promote increased life. However, quicker control could be achieved by placing the sensors near the throttle valve, which is located at the rear, pump end of the system.

For pressure stability a six-stage blower-style dry pump is used with a pressure base almost two orders of magnitude lower than the control set point. This type of pump has an extremely repeatable base and pumping speed, which aid the in situ calibration of the capacitance monitor. The sensor is calibrated using the pump's stable base and its low base pressure. Aided by the AFD system, the high end of the sensor is checked with a set gas flow and throttle-valve position to determine whether proper pressure is reached. This test works only if gas flow is accurate.

Gas Flow. Modern mass-flow controllers (MFCs) are very accurate. However, despite the many advances over the last decade, basic sensor design has not changed. Particles and contamination in the gas-flow stream can lodge in the MFC and cause a calibration change. If handled correctly, MFCs in an ultrapure gas delivery system can last for years without failing or requiring recalibration. Because they are often destroyed by removing them from the system, the ability to check their calibration in situ is critical.⁴

View of the load station of the 6-in. horizontal furnace stack at AMD's Fab 15.

One method for in situ MFC calibration uses a mass-flow meter (MFM) in a common nitrogen purge line to all MFCs, as shown in Figure 2. Valves A and C are used to flow nitrogen gas through the MFM and the first MFC. The proper use of interlocks, gas pressure, and check valves is necessary to prevent the unwanted mixing of gases. With this gas tray system each MFC can be checked against the MFM using nitrogen and each gas's correction factor or curve. The MFM is more reliable than a standard MFC because it flows only nitrogen and is used periodically. Also, MFM drift would show up across all MFCs and thus be detectable.

In designing this type of gas tray, it is important to scale the MFM to match the MFC flows as closely as possible. Flow differences of up to two orders of magnitude can be successful if gas biasing is used. This technique allows small-flow MFCs to be calibrated by first centering the flow of the MFM through an MFC in which the actual flow has already been checked, and then flowing the small-flow MFC through the MFM. Using the center of the MFM increases the accuracy of the checking process for detecting change over time using the AFD system.

Temperature. In almost all chemical reactions, controlling the overriding variable, temperature, is critical. In diffusion and CVD processes, temperatures range from 300° to 1300^°C, making thermocouples necessary. Unlike pressure, which is controlled by means of mechanical standards, and gas flow, which is controlled by means of calibration using MFMs, in situ calibration of thermocouples at high temperatures is difficult or impossible. Temperature electronics can be checked with a calibrated voltage meter, but this does not certify the entire temperature loop. Thermocouples that are calibrated to extremely tight tolerances can be purchased, but drift caused by high-temperature effects causes changes over time.⁵ If the manufacturer calibrates the thermocouples after use, the amount of drift can be statistically quantified for individual applications. The replacement of the thermocouples can then be scheduled in light of temperature drift versus its effect on the process. Temperature measurement will remain reliable if recommended thermocouple measurement practices are followed and thermocouples are replaced periodically.⁵

Gas tray system that uses a mass-flow meter to achieve in situ MFC calibration. Photos courtesy of AMD.

Automatic Fault Detection. Once the proper efforts have been made to ensure that the sensors are producing real, usable data, the next step is using AFD to extract information from the data. AFD is best used to statistically track changes to time-dependent parameters, where alarm/control limits inhibit the start-up of subsequent runs if an out-of-control event occurs. Data from any input, at any time within a run, can be converted to a data point using algorithms such as slope, average, rate of rise, thermal budget (Dt), and process capability index (C^_pk). These data are then plotted over time against other data. Input data can also be plotted for a single run against statistically derived alarm limits, as shown in Figure 3, where throttle-valve position on a low-pressure furnace is plotted versus time.

Figure 2: Schematic of a mass-flow meter in a common nitrogen purge line used for in situ MFC calibration.

Figure 3: Plot of throttle-valve position on a low-pressure furnace versus time.

To control pressure, AFD can be used to track throttle-valve position and furnace rate-of-rise versus run, as illustrated in Figures 4 and 5, respectively. Throttle-valve change is an excellent indicator of deposition buildup in the cold trap, which reduces pumping speed, and in the capacitance manometer line, both of which make maintenance necessary. Furnace rate-of-rise change is an indicator of a system leak, which also requires maintenance. These two parameters are excellent early warning indicators of impending vacuum failures which, if they occur during a run, can ruin hundreds of wafers.

Figure 4: Plot of throttle-valve position versus run used to gauge pressure.

To control gas flow, calibration for each MFC can be tracked over time using the MFM flow results and the gas's calibration factor. Gas flow checks are conducted every 2 weeks as part of an oxide qualification run. Figure 6, which shows actual gas flow for an oxygen MFC, indicates excellent MFC performance over a period of a year and a half.

Figure 5: Plot of furnace rate-of-rise versus run used to gauge pressure.

Figure 6: Actual gas flow versus date for oxygen mass-flow controller.

To control temperature, many different parameters can be tracked. Taking the area under the temperature curve multiplied by time, Dt can be calculated and graphed as an indicator of the amount of thermal energy in a drive process, as shown in Figure 7. Also, power has been tracked during an oxidation step (Figure 8). In that case, average power did not deviate substantially over 500 runs, showing good element performance and thermal packing.

Control System Results

AFD has been running on all furnaces for more than two years with excellent results. When a furnace alarm indicates that an AFD parameter is out of limits, the operator contacts an engineer or technician to diagnose the problem. Most alarms are early indicators that a tool requires maintenance. For many parameters, early warning allows production to schedule downtime so that repairs can be made when convenient. Also, tracking parameters with AFD facilitates preventive maintenance on tools, which can reduce downtime.

Figure 7: Dt as an indicator of the amount of thermal energy in a drive process.

Figure 8: Tracking of power during an oxidation step, indicating that average power did not deviate substantially over 500 runs and that good element performance and thermal packing were maintained.

MFC drift detection has led to a reduction in misdiagnoses of MFC failures, all but eliminating unnecessary gas delivery system repairs and long tool requalifications. When a gas-flow alarm occurs, all the flow data for the system in question are checked to better master the parameter that has gone out of control. Occasionally, all the MFC data drift simultaneously, indicating a general shift in nitrogen pressure or a possible MFM problem. More often a process problem occurs that calls MFC performance into question. In this situation the AFD flow data are used to determine whether there has been a calibration shift.

The early detection of throttle-valve position and rate-of-rise faults in low-pressure furnaces regularly prevents furnace aborts that result in wafer scrap. The information obtained using AFD has been invaluable for troubleshooting equipment and process-related problems. Because pressure is controlled to a set point, the throttle-valve position and deposition rate are the main indicators of a change in actual system pressure. On several occasions a change in the pressure rate-of-rise test has indicated that an O-ring seal has degraded and must be replaced.

On drive processes where oxidation thickness is not a clear indicator of transistor performance, Dt has been calculated and tracked to wafer electrical measurements. The tracking of element power has been implemented. Although no major changes have been detected, it is believed that this parameter is a good indicator of thermal changes in the furnace that might presage element failure or packing problems after a tube change.

Conclusion

The careful addition of only the most necessary functions to AFD has kept the system manageable and running smoothly in production. Large amounts of data are reduced to easily usable graphic information for the most important parameters, which aids in troubleshooting tool problems. AFD has improved the overall performance of the furnaces, increasing tool utilization and reliability. It has reduced wafer scrap by providing early warning of equipment failure, thus saving $100,000 per year and proving itself to be a cost-effective means of defect prevention. Because of their effectiveness, run-to-run controls are up and running for some processes and AFD is being implemented at Fab 25, a state-of-the-art logic fab producing 0.25-µm devices on 8-in. wafers.

Statistical data management has improved the analysis of time-based trends and highlighted areas for further tool improvement. The development of AFD was made possible through a successful partnership between the tool supplier and the end-user, enabling the construction of a full control platform. AFD has underscored the need for better in situ data collection methods.

Acknowledgment

This article is a revised version of a paper originally presented at SemiPAC 99 (San Antonio, TX) in January 1999. It is used with permission.

References

1. M Yelverton et al., "A Complete Furnace Control Platform for High-Volume Manufacturing," in Proceedings of the IEEE International Symposium on Semiconductor Manufacturing (New York: The Institute of Electrical and Electronics Engineers, 1998), 59—62.

2. E Sachs et al., "Process Control System for VLSI Fabrication," IEEE Transactions on Semiconductor Manufacturing 4, no. 2 (1991): 134—143.

3. GG Barna, MM Moslehi, and YJ Lee, "Sensor Needs for IC Manufacturing," Solid State Technology 37, no. 4 (1994): 57—61.

4. J Cestari and M Yelverton, "Maintaining Ultraclean Gas-System Integrity for Toxic and Hazardous Gases," Solid State Technology 38, no. 10 (1995): 109—116.

5. Manual on the Use of Thermocouples in Temperature Measurement, 5th ed. (Baltimore: American Society for Testing and Materials, 1990), 24.

Mark Yelverton is a candidate for member of the technical staff at AMD in Austin, TX. He joined AMD in 1986 as a
diffusion engineer. Partnering with SEMY Engineering (Phoenix), he has been involved in the development of model-based temperature control, automatic fault detection, and run-to-run control for diffusion furnaces. He received his BS in electrical engineering from Texas A&M University (College Station, TX) and his MS in manufacturing systems from National Technological University (Fort Collins, CO), where he focused on modeling and control. (Yelverton can be reached at 512/602-4433 or mark.yelverton@amd.com.)

Brian Cusson is the diffusion module manager for AMD in an 8-in. submicron logic fab. He joined the company in 1989 as a diffusion/metallization yield-enhancement engineer for the contamination-free manufacturing group. After receiving his BS in electrical engineering from the Georgia Institute of Technology (Atlanta), he worked at Harris Government Aerospace Systems Div., VHSIC operations, as a diffusion/implant process engineer. He has been a member of IEEE since 1985. (Cusson can be reached at 512/602-4161 or brian. cusson@amd.com.)

Tom Timmons is an engineering technician in charge of SEMY installations in the diffusion area at AMD in Austin. He has published several papers and obtained a patent for his work in the field of advanced process control. (Timmons can be reached at 512/602-5515 or ttimmons@amd.com.)

Kevin Stoddard is the control systems division manager at SEMY Engineering in Phoenix, which he joined in 1986. He is involved in the development of model-based equipment control, automatic fault detection, and run-to-run process control solutions for the semiconductor industry. He has a BS and MS in electrical engineering from Arizona State University (Tempe). He has written many publications and holds several patents in the areas of advanced process and equipment control. (Stoddard can be reached at 602/861-9395, ext. 229, or kstoddard@semy.com.)

MicroHome | Search | Current Issue | MicroArchives
Buyers Guide | Media Kit