Click here to go to a list of titles for these articles.
Statistical tests are developed to provide an operational definition of the term "under control" for application in deciding whether trendcharts in semiconductor operations should be implemented as control charts, which often require considerable investment in responding to special cause flags. The decision rules derive from Monte Carlo simulations of traditional special cause flags modified for sample size to provide constant and low probabilities of inappropriate process rejection, to help ensure easy tracing of underlying problems and to help ensure large improvements in process capability from investment in fixes. The tests maintain the high efficiency of time series analysis intrinsic to traditional special cause flags. That uncompensated effects of sample size in control chart setup invite wildly uncontrolled risk of rejecting good processes is shown. A table of decision rules is provided to facilitate application of the setup tests at a constant and low risk of rejecting good processes.
This paper considers the run by run control problem. We develop a framework to solve such a problem in a robust fashion. The framework also encompasses the case when the system is subject to delayed measurements. Recent results available for the control of such systems are reviewed, and two examples are presented. The first example is based on the end-pointing problem for a deposition process, and is subject to noise which has both Gaussian and uniform components. The second one is concerned with rate control in an LPCVD reactor.
In microelectronics manufacturing, control strategies for plasma etch systems have been limited to traditional statistical process control and recipe control techniques. The lack of in situ real-time measurements of process performance and appropriate models has hindered the introduction of feedback control systems. This paper focuses on empirical model building for advanced process control using two real-time diagnostic sensors for measurement of the reactor state. Laser interferometry for measurement of etch rate and voltage and current probes for measurement of effective radio-frequency power and sheath voltage, couples with data acquisition hardware and software, provided the foundation for steady-state and dynamic model development of the plasma etch process. Several linear and nonlinear steady-state techniques including ordinary least squares, neural networks, and projection to latent structures were used in empirical model building. Both linear regression and recurrent nerual network model structures provided a satisfactory fit of the data for the operating space investigated. Projection to latent structures techniques indicated that the most relevant variables were power, pressure, and chamber impedance. The addition of the impedance measurement significantly improved the predictive capability of the model.
Everyone is being exposed to the "zero defects" philosophy which establishes zero as a goal. This will not be achieved overnight but approached over time by continually striving to reduce targets. What kind of techniques are needed to assure zero defects? What constitutes an out-of-control situation? An attributes control chart conveys little information at or near zero defects. Assuring zero defects through sampling inspection leads to infinite samples or 100 percent inspection, assuming 100 percent inspection efficiency (the latter rarely exists, and efficiency probably gets worse at lower defect levels). Obviously, some new approaches to quality control (QC) techniques will be necessary at zero defects. One old standy is the variables control chart, OVERLINE X and R, but with the specification at least five standard deviations from the average. Thus one route to zero defects is a properly chosen specification. However if attributes data must be used, the standard p and u charts and not very useful. Perhaps a control chart that plots the number of good items between defects on a logarithmic scale to accommodate large numbers can be used, establishing upper and lower limits on the number of items between defects. Another problem area at zero defects is sampling inspection to assure targets. Following the approach above on good items between defects, the number of accepted lots between rejected lots can be a criterion. Sample sizes can be related to lot sizes as in MIL-STD-105D. If rejected lots come too close together, a procedure can be established requiring process shutdown until the problems are resolved. This will help promote the "making it right the first time" philosophy necessary to achieve zero defects.
In the last few years, "Run-to-Run" (R2R) control techniques have been developed and used to control various semiconductor manufacturing processes. These techniques combine response surface, statistical process control, and feedback control techniques. This paper provides a literature review of R2R control methods from a statistical and control engineering point of view. It is shown that self-tuning controllers can provide a valuable control strategy for R2R applications. In this paper we address the single-input-single-output (SISO) case. Two proposed self-tuning controllers compensate not only for the standard case of process shifts, but also in the case a deterministic trend and/or autocorrelation is present in the observed response (i.e., in case there exist process dynamics). In order to reduce the input variance, a control chart is added to the output and acts as a deadband.
The usual approach to bring the value of the process capability index, Cpk, to an acceptable level isto design equipment and develop a process in which the process variable is robust to external disturbances. However, an alternative approach involves the use of in situ sensors and real-time feedback control. Currently this approach is not widely implemented in the semiconductor industry. In this paper, we provide analytic justification to quantify the potential improvement in Cpk if a real-time feedback control scheme is used instead of the usual open-loop approach. We show that for the case of feedback control the level of Cpk is only limited by the accuracy and reproducibility of the sensor provided that the target values are indeed achievable by the processing equipment. This result also holds in the presence of disturbances and nonlinearities. Cpk values for open-loop and real-time feedback control strategies are compared for two experimental applications: single-wafer CVD nitride and polysilicon processes.
This paper reports on the application of quadrupole mass spectrometry (QMS) sensing to real-time multivariable control of film properties in a plasma-enhanced CVD silicon nitride process. Process variables believed to be most important to film deposition are defined (i.e., disilane pressure, triaminosilane pressure, and dc bias voltage) and their responses to system inputs are modeled experimentally. Then, a real-time controller uses this information to manipulate the process variables and hence film performance in real time during film deposition. The relationships between gas concentrations and film performance are shown explicitly where the controller was used to drive the concentrations to constant setpoints. Also, an experiment investigating the effects of an out-of- calibration mass flow controller demonstrates the compensating ability of the real-time controller. The results indicate that in situ sensor-based control using quadrupole mass spectrometry can significantly assist in optimizing film properties, reducing drift during a run, reducing run-to-run drift, creating a better understanding of the process, and making the system tolerant to disturbances.
The goal of our control system is to improve the reliability, accuracy, and economy of operation of a sequence of interrelated processes. We achieve this task by using well known, rigorous statistical techniques to continuously monitor process parameters, detect out-of-control equipment, and then optimally adjust relevant machine inputs to bring the process back on target. We have implemented the supervisory control system on the photolithography sequence in the Berkeley Microfabrication Laboratory, where it has been conclusively proved that the supervisory control system increase significantly the capability of the entire process. The supervisory control algorithms, consisting of feedback and feed-forward control, multivariate, model-based statistical process control (SPC), and automated specification management algorithms, are independent of machine and/or process, and can be applied to any semiconductor manufacturing sequence.
This paper presents a general diagnostic system that can be applied to semiconductor equipment to assist the operator in finding the causes of decreased machine performance. Based on conventional probability theory, the diagnostic system incorporates both shallow and deep level information. From the observed evidence, and from the conditional probabilities of faults initially supplied by machine experts (and subsequently updated by the system), the unconditional fault probabilities and their bounds are calculated. We have implemented a software version of the diagnostic system, and tested it on real photolithography equipment malfunctions and performance drifts. Initial experimental results are encouraging.
Competition in the semiconductor industry is forcing manufacturers to continuously improve the capability of their equipment. The analysis of real-time sensor data from semiconductor manufacturing equipment presents the opportunity to reduce the cost of ownership of the equipment. Previous work by the authors showed that time series filtering in combination with multivariate analysis techniques can be utilized to perform statistical process control, and thereby generate real-time alarms in the case of equipment malfunction. A more robust version of this fault detection algorithm is presented. The algorithm is implemented through RTSPC, a software utility which collects real-time sensor data from the equipment and generates real-time alarms. Examples of alarm generation using RTSPC on a plasma etcher are presented.
In this paper we propose a profit-based framework for an integrated CADûCAM system for present and future VLSI design and manufacturing. The inefficiencies of present day CAM systems are due to the lack of appropriate methodologies for monitoring lots in fabrication lines using in situ measurements and controlling lots using the multivariate distribution of observable in-process parameters are developed at Carnegie Mellon University. The software system which implements the new algorithms have shown encouraging results when applied to industrial fabrication lines.
In this paper, the first of two parts, we present a methodology for statistical process control of VLSI fabrication processes. We formally introduce a general framework for a computer-aided manufacturing system that can be used to monitor, diagnose, and control IC manufacturing. We formulate the task of process control as one of profit maximization and develop the associated objective function and the constraints for a number of manufacturing scenarios. Finally, we formalize the IC fabrication process as a stochastic system and present the necessary conditions for efficient statistical control of VLSI fabrication processes.
This paper presents the methodology developed for the automatic feedback control of a silicon nitride plasma etch process. The methodology provides an augmented level of control for semiconductor manufacturing processes, to the level that the operator inputs the required process quality characteristics (e.g. etch rate and uniformity values) instead of the desired process conditions (e.g., specific RF power, pressure, gas flows). The optimal equipment settings are determined from previously generated process/equipment models. The control algorithm is driven by the in-situ measurements, using in-line sensors monitoring each wafer. The sensor data is subjected to Statistical Quality Control (SQC) to determine if deviations from the required process observable values can be attributed to noise in the system or are due to a sustained anomalous behavior of the equipment. The updated models are used to run subsequent wafers until a new SQC failure is observed. The algorithms developed have been implemented and tested and are currently being used to control the etching of wafers under standard manufacturing conditions.
This paper presents a process monitoring system, which is designed to be used for monitoring VLSIC and other multistage manufacturing processes. The proposed process monitor can 1) simultaneously defect a variety of out-of-control conditions, 2) quantify the magnitude of process change, and 3) be used to compute the probability of meeting specifications. Average run length simulations show that for a single-stage process, the monitor is at least as good as the Shewhart-CUSUM charts for detecting changes in the distribution of the monitored characteristics. For a multistage process, however, the Bayesian monitor can significantly reduce the detection time by using in-line correlation information from earlier stages. The monitor has been applied to data from a state-of-the art fabrication facility, and the results are promising.
This paper describes the development of real-time control technology for the improvement of manufacturing characteristics of reactive ion etchers. A general control strategy is presented. The principal ideas are to sense key plasma parameters, develop a dynamic input-output model for the subsystem connecting the equipment inputs to the key plasma variables, and design and implement a multivariable control system to control these variables. Experimental results show that this approach to closed-loop control leads to a much more stable etch rate in the presence of a variety of disturbances as compared to current industrial practice.
An adaptive nonlinear controller for wafer-to-wafer plasma etch control is described. It uses real-time process signatures and historical data from a relational database for a computation of the over-etch time for the current wafer etching within the reactor. For an MOS gate etch, the standard deviation of the oxide thickness between the gate and the source (or drain) is in the range of 10A. This is comparable to open-loop control or timed etch where the operator selects the ideal over- etch time. The controller hasthus achieved a minimum of human equivalence and often performs better by 40%.
A number of quality control and yield improvement techniques are used in the Hewlett-Packard (HP) Fort Collins IC wafer fabrication line. Four of these are described, with examples of how each has improved quality and yield. 1) Silicon wafer measurements obtained from the supplier or made at incoming inspection are correlated with device parameters and chip yield. 2) Control charts in manufacturing are generated on-line from monitor wafer data entered by operators, giving immediate feedback. In addition, a daily summary report lists any chart out of control. In certain instances it is necessary to improve the process capability of an operation. A feedback technique is used to do this for operations which have a predictable systematic drift. 3) Individual wafer positions in critical operations are automatically recorded through the fabrication line. This greatly facilitates correlation of input to output parameters and pinpoints the root cause of physical, device parameter, or chip yield fluctuations. 4) Rapid correlation of the myriad of data obtained throughout the fabrication and wafer test areas is done with a common data base and tools which transform the raw data into an optimum form for analysis.
In this paper, the second of two parts, we present the algorithms used to implement the CMU-CAM system described in the first part [13]. We then present the implementation details of this software system. THE CMU-CAM system performs profit maximization through statistical process control. Its capabilities are illustrated by a number of computational examples.
During the last five years, we have witnessed the widespread application of statistical process control in semiconductor manufacturing. As the requirements for process control grow, however, traditional statistical process control applications fall short of their goal. This happens because modern processes are more complex than they used to be. Further, because of the expanding use of the so called "cluster" tools, modern technologies are also less observable than before. Because of these difficulties, we can no longer afford to wait until a malfunction can be detected on a traditional control chart.
Fortunately, modern semiconductor manufacturing tools can communicate to the outside world a number of their internal parameters, such as throttle valve positions, chamber pressures, temperatures, etc. It is intuitively obvious that equipment malfunctions will manifest themselves first in the values of these internal parameters and much later on the water properties. In this paper, we describe a process monitoring scheme that takes advantage of such real-time information in order to generate malfunction alarms. This is accomplished with the application of time-series filtering and multivariate statistical process control. This scheme is capable of generating alarms on true real-time basis, while the water is still in the processing chamber. Several examples are presented with tool data collected from the SECSII port of single-wafer plasma etchers.
Many qualitative properties of the product and the process are of interest during semiconductor manufacturing. One of the typical examples is the sidewall surface roughness of an etched polysilicon line. These properties are important since they affect directly the utility and performance of the integrated circuit (IC) devices being built. Traditionally, however, they are treated informally and subjectively as tacit knowledge in the processing arena. In this paper, we present a systematic approach to modeling and controlling such qualitative properties. This approach is based on treating qualitative process variables as categorical data that can be better understood with the help of formal statistical analysis known as logistic regression. This analysis reveals important relationships between the input process settings and the qualitative process output responses in a way that is similar to linear regression analysis for conventional numerical variables. Similarly, categorical process variables can be used for process control, which is driven by a probabilistic model of the categorical variables. We will show how categorical models can be used to tune a process and, later, to control it via statistical process control (SPC) charts, model-based quality control techniques, and adaptive run-by-run controllers.
An advanced multivariable in-line process control system, which combines traditional statistical process control (SPC) with feedback control, has been applied to the CVD tungsten process on an Applied Materials reactor. The goal of the model-based controller is to compensate for shifts in the process and maintain the wafer- state responses on target. The controller employs measurements made on test wafers to track the process behavior. This is accomplished by using model-based SPC, which compares the measurements with predictions obtained from process models. The process models relate the equipment settings to the wafer-state responses of interest. For CVD tungsten, a physically-based modeling approach was employed based on the reaction rate for the H2 reduction of WF6. The Arrhenius relationship for the kinetic model was linearized so that empirical modeling techniques could be applied. Statistically valid models were derived for deposition rate, film stress, and bulk resistivity using stepwise least-squares regression. On detecting a statistically significant shift in the process, the controller calculates adjustments to the settings to bring the process responses back on target. To achieve this, two additional test wafers are processed at slightly different settings than the current recipe. This local experiment allows the models to be updated to reflect the current process state. The model updates are expressed as multiplicative or additive changes in the process inputs and a change in the model constant. This approach for adaptive control also provides a diagnostic capability regarding the cause of the process shift. The adapted models are used by an optimizer to compute new settings to bring the responses back to target. The optimizer is capable of incrementally entering controllables into the strategy, reflecting the degree to which the engineer desires to manipulate each setting. The capability of the controller to compensate for induced shifts in the CVD tungsten process is demonstrated. Targets for film bulk resistivity and deposition rate were maintained while satisfying constraints on film stress and WF6 conversion efficiency. The ability of the controller to update process models during routine operation is also investigated. The tuned process models better predict the process behavior over time compared to the untuned models and lead to improved process capability.
On-line statistical process control (SPC) has been implemented on a single-wafer remote microwave plasma photoresist asher. SPC for ashing is made more difficult because the prior processes, e.g., ion implantation, affect the properties of the resist material, and consequently the ashing behavior. The system presented comprehends the variety of incoming wafer states from a complex process flow. On-line SPC charts track photoresist clear time on a wafer-to-wafer basis using optical emission spectroscopy. The data is corrected for the "first-wafer" effect, whereby the clear time for a wafer decreases as the delay time between ashing wafers increases. The data is standardized using an expected time and variance for each process flow level to allow all results to be presented in a single set of individuals and moving-standard deviation Shewhart charts. Standard SPC rules are applied automatically within each process flow level to test for unnatural variation in the data. Observed abnormal behavior is due mainly to changes in the incoming material for a specific process flow level, not deviations in the ashing process. When a shift in incoming wafer state is detected, the expected response for that process level is automatically updated to reflect the change. The usefulness of one-line monitoring as a means for identifying misprocessing is prior process steps has been demonstrated. Early diagnosis can save money by avoiding expensive downstream processing on previously misprocessed wafers. In our demonstration laboratory, the equipment has processed wafers from a dozen process flow levels.
No abstract available.
Back to Home Page.