Advanced Process Control Systems:

The Next Step in IC Manufacturing

To reduce manufacturing costs and prepare for future developments in wafer size and line width, many semiconductor manufacturers are looking to advanced process control systems. Sematech supports this trend and the SIA Semiconductor Roadmap calls for significant advances in this area.

Advanced control systems use existing or newly developed sensors to determine conditions in the processing tool. Depending on its sophistication and the benefit it can provide, a controller may use the sensor data to perform a number of different functions, including fault detection, run-to-run control, and real-time control

Fault Detection

The most rudimentary function of these new control systems is fault detection. In fault detection, the sensor detects a condition that is out of preset specifications and the controller prevents further processing until the fault is fixed. Detecting a vacuum chamber leak and preventing deposition or etch of the next wafer is an example of a fault detection technique. This type of control is highly effective in preventing scrap.

Texas Instruments (TI) has had a real-time fault detection control system installed in a dry etch system for a number of years. The controller monitors tool conditions and prevents processing if the conditions go out of specification. A detailed study conducted last year identified ten faults that the controller identified and prevented from causing further harm.

      The first fault was a chamber leak. The effects of the leak were showing up intermittently, and it was only through the use of a continuous monitor that the problem was identified. In this case, the controller prevented processing of twelve wafers that were queued for the leaky tool. Because the problem was intermittent, the engineer concluded that the controller may have prevented many more wafers from reaching the scrap pile.

      The second fault detected was the selection of an incorrect process recipe. The actual fault detected was that the process reached an early endpoint, so the controller halted the process prior to overetch. Unfortunately, the wafers were still processed with the incorrect recipe, so this type of control was not the ultimate answer to the fault. TI has since implemented automatic recipe download, which eliminates this type of fault.

      The third fault was poor wafer placement on the electrode. In this case, the controller identified the error and halted processing for nine wafers during the study period, preventing misprocessing of all nine. The controller also detected and prevented processing when the chamber was improperly cleaned, when arcing occurred, and when the RF power shut off.

Throughout the study, the researchers estimate that the controller saved at least 160 wafers from misprocessing. Unfortunately, the controller also detected faults that did not involve a jeopardy situation for the material. In these cases, the controller cost money. Specifically, the controller caused lost production time and required otherwise unnecessary maintenance and engineering resources.

In spite of these losses, the controller lowered overall production costs enough for TI to use it in their production facility. At the same time, they have been working diligently to eliminate the false faults.

Run-to-Run Control

Run-to-run control takes input from metrology tools (either sensors within the tool or external instruments) and adjusts tool parameters based on the results obtained on the last run. Like fault detection, run-to-run control also reduces scrap, but, more importantly, actively reduces run-to-run variation.

Adventa Control Systems markets a run-to-run controller that has been in use at TI since 1993.

The controller incorporates a process model and control strategy defined by the responsible process engineer. With the user-defined control strategy defining the control algorithm, the controller tunes the process model based on empirical data. The improvements in run-to-run repeatability have enabled users to reduce the number of test wafers, and, concurrently, increase the period of time between maintenance procedures.

Although the etch process is one of the most complex processes in semiconductor manufacturing, simply adjusting etch time based on process data can significantly improve the process. In several installations, the controller takes data from sensors inside the process tool (mass flow controllers, pressure transducers, etc.) and the thickness measurement from the previous run, and uses this data in the process model to calculate a new etch time. Figure 1 shows that this technique compensates for a nearly exponential drop in etch rate and keeps the process within very tight thickness specifications.

Data from more than 400 installations show the Adventa system increases machine availability by 10% to 20% and reduces test wafer usage by 10% to 70%, resulting in an overall fab capacity increase of 12% to 17%. If reduced test wafer usage doesn't justify installation of an advanced control system, increased probe yield may: installation of the Adventa controller has increased Cpk by up to 70%. By tightening the control window, users have obtained probe yield increases of 2% to 5%.

Real-time Control

Even Carl Fiorletta, Marketing Director for Adventa, agrees that run-to-run control is an interim step in semiconductor process control. The ultimate control system looks at actual wafer and process conditions and adjusts the tool parameters while the process is running. To date, this technique has been applied primarily to single variables in relatively long processes such as oxidation and LPCVD.

SEMY Engineering has incorporated Model-Based Temperature control into its vertical furnace controllers. The controller uses both spike thermocouples (thermocouples external to the process chamber) and profile thermocouples (thermocouples inside the process chamber, close to the wafer load) along with power delivered as input to the control algorithm. Using a dynamic, empirically-derived multi-variable model, the controller then calculates a new set point for the power to the furnace heating element.

Although this controller is looking at only one variable in a multi-variable process, the results of using this technique have proven to be significant. Using real-time control of temperature on five different oxidation processes has resulted in significant Cpk increases (see Figure 2).

Voyan Technologies takes real-time temperature control one step further. Empirical studies of vertical furnaces revealed that the edge of a wafer responds to temperature changes more rapidly than the center. This phenomenon causes the wafer edge to be hotter than the center during ramp up and cooler than the center during ramp down. For processes with within-wafer non-uniformity, the Voyan controller uses a novel technique to even out the non-uniformity: time-varying set points. The Voyan controller gradually heats and cools the wafers to create a uniform process. With limited data to date, this control strategy has already shown throughput improvements of 33% and within-wafer process improvements of 28% (the thermal oxide film thickness improved from a range of 0.70 to 0.46 ).

These examples may seem conservative compared with what Xylaur Enterprises has in mind. Xylaur plans to analyze the intermediate species of a chemical vapor deposition (CVD) or etch process in real-time and use these data to fine tune process set points, such as gas flow, pressure, and temperature, to keep the reaction chemistry in the boundary layer above the substrate constant. For CVD reactions, the theory behind this type of control is that constant reaction conditions in the boundary layer (i.e., partial pressures of all gas phase components, including transient species, temperature, etc.) yield constant thin film properties (i.e., thickness, refractive index, resistivity, etc). The same theory would apply run-to-run or even chamber to chamber.

Xylaur has demonstrated the ability to sense the intermediate species real-time in a TEOS deposition process and has begun to correlate the sensor data with the resulting film properties. They expect the control portion of the project to follow on quickly.

Difficulties in Implementation

Brad van Eck, Sensor Integration Project Manager for Sematech, points out that the biggest difficulty in a control scheme like Xylaur proposes is not in the technical difficulty of the project, but in the reluctance of the equipment manufacturers to incorporate both the sensors and the control scheme into their hardware. Unless the equipment has ports for sensors and communications capability, sensor integration is nearly impossible.

For the equipment manufacturers to buy into advanced process control, they need to hear about the need for it from their customers. Unfortunately, many process engineers are not ready to accept that a changing set point can offer improved process uniformity. Many control engineers working in IC manufacturing houses lament that the only 'knob' the process engineers will allow them to control is the process time.

But this is not a new problem. When mass flow controllers (MFCs) were first introduced, it was thought they would be used only for R&D  they would be great if you wanted to change your set point frequently, but no one would want to sacrifice the repeatability of a fixed rotameter for an expensive, potentially unreliable MFC. Yet, as soon as the engineers saw the process improvements afforded by the MFC, they adopted it at once. The equipment manufacturers reluctantly changed from rotameters to MFCs, and within a few years MFCs were standard.

From the experience of Adventa and other pioneers in the area of advanced process control, it is apparent that sensors and advanced control strategies will also be adopted with adequate proof of process improvement.

A second hurdle for advanced process control is communication between system components. Sensors range from thermocouples, providing single-point data in millivolts, to residual gas analyzers, providing data on up to 300 channels through a customized interface. Sematech has attempted to provide some leadership in this direction, publishing a specification for bus-based communications called Sensor Bus. The specification called out three 'acceptable' network technologies: Allen Bradley's DeviceNet, LonWorks, and Honeywell's SDS. DeviceNet appears to be the network with the most support at this time (Applied Materials and several IC manufacturers have standardized on DeviceNet).

Development of extensive communications capability such as DeviceNet may not, however, be cost effective for manufacturers of simple sensors. So while DeviceNet has taken an early lead as the preferred approach, it is not clear that any common communications standard will ultimately enable rapid development of advanced process control technology. In fact, Adventa has taken a communications-independent approach with their run-to-run controller.

Adventa uses a 'black box' communications box for data input. This versatile unit accepts data in SECS/GEM formats so it can accept data from the tool controller, and has a series of analog and digital I/O ports for direct input from many different types of sensors. The black box can make the raw sensor data available on the network or it can process the data before transmission. Communication of new set points back to the tool controller is also provided via the black box, providing another crucial part of the controller's communications needs. Adventa also allows an operator to enter data through a keyboard, adding another level of flexibility to the system.

Some groups have proposed a more pervasive communications network: Ethernet. Although Ethernet requires a higher level of communication than Adventa requires, Ethernet offers some distinct advantages, especially if smart sensors are used:

The sheer volume of sales for Ethernet components keeps hardware costs low. In-house MIS groups or other readily available experts can maintain the network, reducing maintenance costs.

Sematech's van Eck feels the push toward Ethernet may be dragging the sensor manufacturers in too many directions. Many sensor manufacturers have just completed programs to integrate DeviceNet. They would need to complete similar programs to integrate Ethernet, and with current cash flow conditions in the industry, this may not be a welcome move. If, on the other hand, industries outside the semiconductor industry are adopting Ethernet, sensor manufacturers will gain a broader marketing base, and semiconductor engineers will gain a larger base of 'plug-and-play' sensors from which to draw.

Additional graphics.

Conclusion

Communications between control systems and sensors is a crucial area of advanced process control, but is one that may not be resolved right away.

Sematech, SEMI/Sematech, and the Integrated Measurement Association recently held its tenth symposium on Advanced Equipment Control and Advanced Process Control (AEC/APC). The more than 50 presenters showed significant progress in all areas of process control, from novel sensors to sophisticated control systems. The process results that were presented showed the potential for significant improvements in both equipment effectiveness and process yield through advanced control techniques.

With results like that, it won't be long before many more IC manufacturers adopt advanced process control techniques.

 

References

1 The Semiconductor Industry Association National Technology Road-map for Semiconductors, 1997 edition, p. 183.

2 Rob Winkler, 'Clearing Up the Confusion on Ethernet,' Sensors, Jan.  1999, pp. 28-35.

A member of SEMI and the Semiconductor Safety Association,  Lise Laurin, a former Intel engineer,  is founder/president of ClearTech, a technical marketing communications firm. Contact her at 14 S. Main St., Newton, NH 03858; 603-382-7682, fax 603-382-0491; llaurin@greennet.net