Digital Mass Flow Controllers come of age
11/01/1996
Digital mass flow controllers come of age
Charles F. Drexel, DXL, Torrance, California
Digital mass flow controllers (DMFCs) have been widely used in production environments. Their unique construction has solved common process and reliability problems, such as accuracy, drift, unwarranted swapouts, surrogate gas calibration, high cost of ownership, and control over a wide range of flows.
Since they were introduced in 1991, DMFCs have accumulated more than two million hours of operation in production tools, usually in the most difficult applications. Under these conditions, DMFCs have demonstrated a mean time between failures (MTBF) of 895,000 hours, 25 times the average for analog controllers. This paper discusses how digital controllers deal with some of the common problems encountered in the fab.
Externally, they look like their analog MFC cousins, fitting into the same plumbing and electrical connectors. They sense mass flow from the temperature difference between two temperature sensors in thermal contact with the gas stream and then process the information digitally with a microcontroller. DMFCs were designed to directly replace MFCs, communicating in both digital and analog modes, since it may be some time before tools are available to connect directly with digital components. An interface card between the tool harness and the DMFC provides connection to a PC via an RS232 line for access to the features provided by the microcontroller (Fig. 1). The first DMFC design was aimed at a number of problems and complaints voiced by process and equipment engineers in an industry survey.
 |
Figure 1. The DMFC provides connection to a computer via an RS232 line to access the features provided by microcontroller.
Accuracy and response at low flows
As the average size of semiconductor features shrinks, the films become thinner, requiring shorter processing time or lower gas flow rates, or both. Process times of =10 sec and gas flows of 3-5 sccm are becoming commonplace. Furthermore, the acceptable tolerance on flow rate accuracy is decreasing beyond ?(1-5)% of the full scale. Such stringent requirements will soon be the norm and demands on flow controllers will be more severe as the industry moves into 0.25-?m geometry and beyond.
At flows below 10-20 sccm, conventional analog MFCs have three weaknesses: response, flexibility of range, and accuracy. This part of the flow range is difficult because analog circuits use fixed components (resistors and capacitors) for tuning, confining them to one specific point on the performance map. As an example, it is traditional to specify response as the time to settle within 1 or 2% of final reading after a 100% step change. In most processes, however, crisp response is required at one or more intermediate flows. If an MFC is tuned for a one-second response to a 100% step change, flow stabilization may require a minute or two for a step from 0 to, say, 10%.
 |
Figure 2. System diagram for a DFMC. The microprocessor computes the final control valve voltage and solves the response (PID) equation.
DMFCs can provide quick response at all flows as a result of digital signal processing. The analog sensor output is amplified and digitized before it is sent to a microprocessor to compute final control valve position. Signals are sent in for desired response and linearization of the output (Fig. 2). Algorithms can be written to be as versatile as necessary to optimize the computations. Proportional-integral-derivative (PID) control coefficients can be designed to allow for variables - such as sensor response, valve response, flow rate, magnitude of step flow changes, temperature variations, and pressure spikes - allowing controller accuracy and response to remain essentially the same over the entire usable flow range.
Examples of DMFCs in low flow applications:
An experimental tungsten deposition process required very low flows of WF6. A DMFC with 2.0 sccm full scale flow was calibrated over a range of 0.04-2.0 sccm (50:1). The response to 20% flow steps over the entire range was < 2 sec and the flow accuracy was within ?2% of the setpoint. Although this process was never used in production, it illustrates an extreme condition that a digital controller can meet.
A new etching process for III-V materials required flows between 2 and 50 sccm. The development engineer determined from tests that DMFCs were the only practical means to meet the specification. A number of custom etchers built with DMFCs have been in satisfactory service for three years.
Wide flow range
Many new etch and CVD processes require two or more sequential flow settings (for example, 100% to 5% or vice versa). If flows are not accurate or if the settling time between flow changes exceeds a few seconds, the results can be unacceptable. In an etch process, if the mix ratio between gases changes momentarily during a step change in flow rate, gas phase reactions may cause a haze to appear on the wafers. Analog MFCs have a specified flow range of 2-100%, but manufacturers recommend use close to the full scale flow [1]. This limitation, partly the result of having to use fixed components for tuning, means two MFCs and associated plumbing are needed to cover two-step processes. DMFCs, on the other hand, can control crisply over the entire 50:1 range.
Examples of wide flow range processes:
A tungsten CVD process using WF6 as the tungsten source requires the deposition of nucleation sites followed by deposition of the film. The first step lasts only a few seconds at a WF6 flow of 4-5 sccm, while the second is at 100 sccm. Flow accuracy is critical to film uniformity. The process is normally run using two analog controllers, one for each flow. A fab having yield problems decided to try using one DMFC instead and observed immediate yield improvement. As a result, all tools were retrofitted and have been operating successfully for two years. As an unexpected bonus, the DMFCs have never needed service, while the MFCs they replaced had a lifetime of 3-8 months.
An etch process required gas flow at 90 sccm, followed by a 5 sccm stream. If the change from one to the other took more than 3 sec, the transient variation in mixture ratio between the gases would spoil the wafer. Two MFCs are usually used, but using a DMFC with only one controller prevented damage to the wafers.
Accuracy
The third performance parameter, accuracy, is also easier to achieve with digital signal processing because of the flexibility of programming algorithms. The accuracy of DMFCs is rated at ?1% of setpoint, while analog MFCs are rated as ?1% of full scale. Thus, the percentage accuracy of an analog controller with respect to the setpoint is simply the ratio of the full scale to the setpoint. For example, a 100 sccm analog controller at 20 sccm will have a rated accuracy of ?5.0% of the setpoint compared to ?1.0% for the digital.
Example:
In an etch process using BCl3, the etch rate is quite sensitive to gas flow rate. Variability of 10% is not uncommon and is thought to be caused by gas flow variations of ?(3-6)% of the setpoint. The use of a DMFC with flow stability of ?1% of the setpoint was able to reduce etch rate variation to ?5% and provide longer controller life.
Surrogate gas calibration
 |
Figure 3. Comparison of the linearity of analog and digital controllers. The output of DFMCs is linearly proportional to the relative mass flow over the measured flow range, whereas the MFC outputs are only linear in the first 33% of the tested range.
A long-standing problem with MFCs has been poor accuracy when calibrating with a surrogate gas instead of the process gas. This procedure is often necessary since more than half of the process gases are hazardous and unsafe for use in an MFC manufacturing plant [1]. The accuracy problem is caused by nonlinearity of the measuring section, heat transfer characteristics of the sensor, relative inflexibility of analog circuitry, and large differences between physical properties of the process gas and surrogate gas. The primary physical differences are gas specific heat and density. The former affects the heat transfer between the sensors and process gas, and the latter affects sizing of the valve and bypass. A secondary difference is due to thermal conductivity. To partially overcome these drawbacks, analog controllers are often tested with a benign surrogate gas that has physical properties similar to those of the actual process gas. Digital controllers, on the other hand, can be switched from one gas to another with minimum error, yet they are calibrated exclusively with nitrogen (Fig. 3).
Example:
A test by SEMATECH [5] compared a digital DMFC with two popular MFCs, all calibrated with a surrogate gas and later tested on a flow bench with the actual process gas. Gases such as CF4 were used by the MFC manufacturers as surrogates for the actual gas, dichlorosilane (SiH2Cl2). Bench testing of the analogs across the flow range gave errors of ?14% and ?15%, while the digital controller, calibrated with nitrogen, had an error of only ?0.5%.
Drift
One of the major weaknesses of analog MFCs is long-term drift. Some large chip producers remove all the MFCs in the fab once or twice a year and send them back to their manufacturers for recalibration, replacing them with a set of spare controllers. Tools for vacuum processes usually incorporate built-in rate-of-rise (ROR) flow calibrators to partially skirt this problem by periodically checking calibration and offsetting the process controller to compensate for drifts. This design assumes that the entire calibration curve has shifted (zero drift) and does not account for a change in circuit gain.
Drift occurs because most of the terms in the mass flow equation that are assumed to be constant are actually variable. In a conventional mass flow sensor, two resistance thermometers located close together on the outside of a capillary tube (standard configuration for all analog and digital mass flow controllers) are used. The mass flow rate can be generalized as a function of the temperature difference and other thermodynamic variables as shown:
 |
where
m? = mass flow rate
dT = temperature difference between two resistance thermometers
Cp = gas specific heat at constant pressure
k = thermal conductivity of material surrounding sensor
h = convective heat transfer coefficient
t = thickness of insulation between thermometer and tubing
T1 = ambient temperature
T2 = gas temperature
R1 = resistance of first thermometer at standard temperature
R2= resistance of second thermometer at standard temperature
G = gain of amplifier
(m? and f? refer to the first derivative with respect to time)
The parameters in the function, f? , are always assumed to remain constant, but that is not really the case. Some change with ambient temperature, some with gas temperature, and some with time. Even though manufacturers try to minimize changes by careful design, stringent quality control, and long burn-in time, calibration is still an open loop process.
 |
Figure 4. The data shows various malfunction categories of MFCs in a fab and the overall failure rate in each group. Tests show that as many as 50% of the removed controllers were misdiagnosed and replacement was unwarranted.
A DMFC can use the microprocessor to provide closed loop control of calibration accuracy. For example, after the controller is calibrated at the factory, the output of the controller to a sensor stimulation can be stored in memory when there is no gas flow. The stimulation mimics a flow signal by creating a false temperature difference and exercising all parts of the system that contribute to the readout. When the same stimulation is given later at zero flow, the drift is then proportional to any existing difference between the output and the stored value. The DMFC returns itself to its original accuracy by adjusting its calibration coefficients. The whole process, known as Autocal, takes about 90 sec and can be enabled to occur manually or automatically. Since Autocal is installed in the DMFC instead of the tool, it can be used throughout the fab. In actual applications where data has been recorded for a year or more, the drift has stayed within 0.15%.
Examples of the DMFC controlling drift:
Using Autocal, the etchers for III-V materials mentioned earlier have consistently kept the controllers within factory calibration for two years.
DMFCs installed in diffusion furnaces often have Autocal enabled as an aid in diagnosing suspected DMFC failures (diffusion furnaces do not have rate-of-rise calibrators). Experience has shown that this method quickly determines whether a shift in flow rate is the cause of the problem.
The procedure Autocal tests for and corrects a shift in gain using the slope of the calibration curve. Mass flow controllers can also experience a zero shift that changes all readings over the flow range by the same amount. Some analog controllers have a built-in provision to test for zero shift, but it is rarely used because there is no certain way to know that the flow is actually zero when the check is made. DMFCs, on the other hand, have an "Auto Zero" algorithm that checks to see that there is no sensor flow before performing the "Auto Zero" function.
Diagnostics
As the price for new fabs soars over two billion dollars, the cost of MFC replacement becomes a major concern. A recent cost of ownership study [2] surveying 7500 MFCs and DMFCs in five complete fabs and two production departments showed that the average annual cost of ownership of an analog MFC is over $15,000, while that of a DMFC is only $1200. In one large fab with 4000 MFCs, the difference added up to $55 million/year. Figure 4 shows the reasons reported for replacement. Half of the controllers removed functioned properly when later tested on a bench, confirming a survey by SEMATECH and highlighting a substantial advantage of digital circuitry when performing diagnostics.
The DMFCs in the study were able to identify the nature of malfunctions and confine replacement to cases where the DMFC was the cause by accessing a large number of its internal parameters and devising tests that could detect problems in the gas system. Additionally, the diagnostics are combined with modem support, using a computer emulator program, so that DMFC factory engineers can pinpoint and correct problems by phone.
The cost-of-ownership study did not specifically address tool productivity caused by mass flow controller reliability; however, it is apparent from the data that tool downtime due to MFC replacement is significant. The reliability of tools is receiving increased attention as a major cost of semiconductor manufacturing [4, 5].
Example of diagnostics:
In the survey mentioned above, one fab included is exclusively equipped with DMFCs. In 1994, the year surveyed, not a single controller was mistakenly removed. This performance was attributed to the diagnostics feature and factory online support.
Replacement spares
A second major cost was incurred by lack of a spare controller when a main controller needed replacement. It is traditional to purchase one or more spares for each rating of controller in the fab, but it is also common that a spare is not available when required. The hourly depreciation cost of modern tools while sitting idle waiting for a replacement MFC is a large part of the high cost of ownership. This emphasizes the most obvious advantage of DMFCs: they can store multiple calibrations on board and be recalibrated in situ, with the gas turned off, using a computer. Experience has shown that DMFC-equipped fabs need 20-30% fewer spares than normally required.
Digital communication standard
The introduction of digital controls illustrated the need to adopt a fab standard for digital communication between tools and computers. SEMATECH set up a committee to solve this problem but it was unable to reach a consensus and eventually abandoned the effort. SEMI picked up the task and also gave up after two years. It therefore appears that a standard will not be selected until digital components are in general usage. At that time, the flow controller manufacturers must modify software and firmware to accommodate the chip set and protocol chosen. In the meantime, the considerable benefits of digital flow control do not depend on adoption of a universal standard.
Conclusion
Digital mass flow controllers have an extensive record of solving difficult flow problems. They can be directly interchanged with analog controllers without electrical or mechanical modifications. A laptop computer may be used for installation, maintenance, calibration, change of gases, diagnostics, and remote support, while the tool process controller handles recipes. Replacement of existing analog controllers with DMFCs can be justified by a short pay-back period and improved wafer quality.n
References
1. George Chizinsky, "Recent Advances in Mass Flow Control," Solid State Technology, Vol. 37, Sept. 1994.
2. Charles F. Drexel, "Cost of Ownership of Mass Flow Controllers," unpublished, available on request from DXL, 2540 W. 240th St., Torrance, CA 90505.
3. Robert Ristelhuber, "Wafer Fabs: Getting More Bang for the Buck," Electronic Business Today, July 1996.
4. Brad Mattson, "Rising to the Challenge of Cost of Ownership," Electronic Business Today, July 1996.
5. David Nelson, "Nonlinear Mass Flow Controller Response Phenomenon," SEMATECH Non-Confidential Report, 91100730A-TR, Oct. 25, 1991.
CHARLES F. DREXEL received his BSME from the University of Colorado and his ME degree from UCLA. He formed DXL in 1988 to develop a new generation of digital mass flow controllers, and has formed two other companies in the US, three in Europe, and one in Japan. Drexel has served as CEO and chairman of Tylan Corp. and holds 20 patents. In 1987, he received the SEMMY award from SEMI for invention and
commercialization of the mass flow controller. He also served two terms as chairman of the Southern California Export Council, three terms as a director of SEMI, and one term as its president. DXL, 2540 W. 237th St., Torrance, CA 90505; ph 310/784-5455, fax 310/784-5464.