Achieving process control through direct measurement of plasmas
11/01/2001
John Scanlan, Scientific Systems USA, San Jose, California
Chris Bowker, LSI Logic Inc., Gresham, Oregon
overview
Measuring and baselining plasma parameters via automated, software-controlled plasma system fingerprinting, instead of relying on best-known methods, provides effective in situ process measurement and control that offers rapid return on investment. The value of this work has been proven in production operations to detect and correct etch rate failure problems on an etch tool during standard production runs, among other examples.
Despite the trend in wafer processing toward advanced process control, in most wafer fabs process engineers still handle process tool fault detection and classification using existing "best-known methods." While logic indicates that if process inputs are controlled, a process will be in control, fab engineers continually report changes in processed wafers, but see no apparent change in process inputs. Or, they find that process inputs do change, but still must determine, classify, and correct the fault.
Overall, fab engineers continue to rely on secondary or machine-state parameters for process control. The industry's increasing demands for rapid ramp rates, increased throughput and yield, and low, nonproduct wafer use are putting stress on these approaches. There is clearly a need for a more rapid and direct approach to process control and fault classification.
In tool fault detection and classification, the objective is to reduce unscheduled downtime by specifically identifying the core problem. With the cost of unscheduled downtime often exceeding $10,000/hr, a process engineer is constantly seeking ways to identify, classify, and correct these faults as rapidly as possible.
Process measurement
The fundamental problem with current methods of process control is that they are typically based on process inputs, such as gas flows, reactor pressure, and generator power. The premise is that if these secondary parameters are in control, process outputs such as etch rate and uniformity will be in control.
In plasma-enhanced processes, however, wafer processing is a more direct function of plasma events than of process inputs. This fact is especially significant when we consider that approximately two-thirds of all process steps in IC manufacturing are plasma-enhanced processes. While the secondary parameters defined do affect the plasma significantly, they represent only some plasma inputs; even the best measurement and control equipment cannot measure all inputs.
 Figure 1. A typical baseline impedance fingerprint of a process plasma showing 15 rf parameters (five voltage, five current, and five phase). |
If we could measure and control the plasma itself, we would not need to be concerned about inaccurate or incomplete methods of measuring a very large combination of inputs. Measuring secondary parameters remains important, but represents an incomplete picture. The answer is in situ process measurement and control in which the plasma itself is measured.
In our work, we have used a nonintrusive rf sensor called SmartPIM. This sensor simultaneously measures the first five Fourier components (harmonics) of current, voltage, and phase, thereby providing an electrical fingerprint (i.e., an impedance fingerprint or ImPrint) of a process chamber (Fig. 1). The rf sensor measures at the chamber on the plasma side of the match unit.
Each fingerprint is unique and stable for a given process recipe; run-to-run repeatability is generally <1%. Any changes in plasma conditions, through process input changes or chamber conditions, result in a fingerprint change. This change can be used to determine and classify the root cause of the shift, ultimately isolating hardware and process problems and identifying changes in individual process inputs.
The fingerprint process
The ImPrint software allows analysis of the impedance fingerprint to provide four functions:
- process chamber matching — comparing real and imaginary impedance of several chambers, and comparing process plasma, inert plasma, and plasma-less chamber conditions;
- process chamber qualification after preventive maintenance — comparing a fingerprint to an archived baseline fingerprint to identify changes where significant variances indicate a shift in plasma or chamber properties that may result in scrapped production wafers;
- process chamber fault classification — helping process engineers rapidly identify a fault condition or eliminate a set of possible failure modes; and
- process chamber fault diagnostics — tracing faults and categorizing them as hardware (e.g., a misplaced focus ring), chemistry (e.g., a faulty MFC), or plasma (e.g., an rf fault).
So, for fault classification, the ImPrint software allows users to easily run a design of experiments on selected process inputs around a given baseline to learn trends on all 15 parameters. The response of rf parameters to each process input is unique. If a fault occurs on a given process chamber, a user can compare fault and baseline fingerprints. The result — the output from the ImPrint software — is a Pareto chart indicating a statistical measure of how close the fault is to any of the learned trends from the selected process inputs.
 Figure 2. Fault diagnostics logic tree based on, as an example, a change in etch rate during production, following PM or when commissioning a new chamber. |
Of course, faults not identified through the design of experiments do occur. For example, hardware may fail or be installed incorrectly following preventative maintenance (PM). Here, fault diagnostics (Fig. 2) allow a process engineer to trace and categorize faults. Categorization can be very effective in reducing unscheduled downtime, and can be extended to allow an engineer to solve chamber-to-chamber differences and set post-PM checks.
Since implementing SmartPIM fingerprinting capability at its Gresham, OR, facility in early 2001, LSI Logic has now installed it on many tools, including Lam Research 4520Xle and 9400 etchers. As a result, fab engineers there have experienced rapid and significant benefit, notably reducing mean time to repair and unscheduled downtime, and speeding chamber comparison and matching. Estimates show that the return on investment for SmartPIM with ImPrint use is 78% after the first year of use and >200% at the end of the second year.
Chamber matching
Chamber-to-chamber differences are a common problem in many fabs. For example, at LSI Logic, one particular chamber difference manifested itself as an "rf abort" fail during a process etch step on a Lam Resarch 4520Xle chamber. The problem meant that this tool would frequently abort during production, resulting in manual interjection and scrapping of a product wafer. This fault often occurred several times a day.
Engineers generated an impedance fingerprint of the problem chamber and three other similar chambers in the fab, generating real, imaginary, and vector impedance values for a standard production process recipe. Impedance Z = R + jX, where R is the real resistive component impedance and X is the reactive (imaginary) component. Vector impedance
 |
null
The problem chamber (i.e., Chamber 2 in Fig. 3) showed lower impedance relative to the others, meaning the match unit experienced sporadic problems in tuning to this impedance, thus the rf aborts.
 Figure 3. Chamber impedance comparison. |
With impedance identified as the problem, process engineers used chamber matching to determine the plasma-less impedance of each chamber. Noting a similar large difference, this indicated that the problem was hardware, not plasma or chemistry. Next, they examined the circuitry and design of the problem chamber, quickly finding (in a few hours) that the impedance difference was likely due to a fault in an input filter board on the power delivery line. Once this board was replaced, the problem chamber's impedance was much closer to the line average and the rf abort problem disappeared.
For this specific example of chamber matching, LSI Logic conservatively estimates savings/chamber/quarter of more than $2000.
Qualification after PM
Process engineers at LSI Logic have also used plasma fingerprinting to detect post-PM catastrophic failures caused by changes to hardware or recipe configurations or by the use of defective parts. Standard post-PM qualification checks, performed on test wafers, may not always pick up these problems. The possibility of significant yield loss exists unless these problems are discovered in some way before electrical test or sort. The software compares the current post-PM fingerprint to archived post-PM fingerprints. Its use means greater sensitivity to hardware and software configurations, and therefore the ability to isolate major post-PM problems.
Following any PM, LSI Logic engineers compare a post-PM fingerprint to its baseline. This method can find the installation of, for example, an incorrect focus ring that will impact device yield through higher defect density because of increased particles. The estimate of savings for this type of failure is ~$9000/chamber/quarter.
Process chamber fault classification
At LSI Logic, process engineers were also experiencing etch rate failure problems on an etch tool during standard production runs. Instead of best-known methods (i.e., checking all tool parameters that affect etch rate, including etch gas flow, the rf generator, the match unit, pressure, and spacing), we used fingerprinting to an established baseline. Trends in some rf parameters pointed toward a change in pressure. In this case, an engineer determined that a pressure manometer was defective. Looking at improvement in mean time to repair this fault, we determined a savings of ~$10,000/chamber/quarter.
We found that a major advantage of our fingerprinting method is that it is sensitive to electrical anomalies in process chambers and to gas pressures and chemistries. At LSI Logic, fingerprints of small air leaks are learned, and then included in the portfolio of fault classifications. The fingerprint incorporates many of the advantages typical of optical and mass spectroscopic tools.
Conclusion
Increasing ramp rates and shorter process transfer times require improved methods for chamber and process matching in plasma processing. Furthermore, reducing unscheduled tool downtime and nonproduct wafer use requires advanced fault detection and classification techniques. With these requirements, semiconductor manufacturers and equipment suppliers cannot depend on ex situ process control and machine-state parameters to meet their needs. Measuring and baselining plasma parameters via automated, software-controlled plasma system fingerprinting provides effective in situ process measurement and control for rapid ROI.
Acknowledgments
Major contributors to the content of this work and article include Justin Sato, plasma etch process engineer, and Peter McGrath, staff process etch engineer at LSI Logic; and Justin Lawler, plasma process engineer, Ciaran O Morain, VP, and Alan Hynes, product manager at Scientific Systems in Dublin, Ireland. SmartPIM and ImPrint are trademarks of Scientific Systems Ltd.
John Scanlan is chief technologist for Scientific Systems USA, 111 N. Market St., Suite 621, San Jose, CA 95113; ph 408/995-5974, fax 408/351-3623, e-mail [email protected].
Chris Bowker received his BS from the University of Colorado. Bowker is plasma etch engineering manager for LSI Logic Inc., Gresham, OR.