Papers

AVS1996 Session MS-MoA: Sensor-Based Fault Detection and Process Control (Continued)

Monday, October 14, 1996 1:30 PM in Room 201A

Monday Afternoon

Start	Invited?	Item
1:30 PM		MS-MoA-1 Fault Detection and Process Control Projects at Sematech B. Van Eck (Sematech) Semiconductor wafer manufacturing is driven largely by customer demands and profits. The desire to maintain profitability motivates the semiconductor industry to continuously improve both product performance and manufacturing efficiency. These improvements typically include increasing device speed and decreasing the cost per function. These improvements require decreased device dimensions, increased wafer diameters, increased yields, and improved equipment productivity. Of these, equipment productivity and device yield have the most direct impact on cost and profitability. As devices shrink and the number of levels of interconnect increase, maintaining competitive yields and costs is becoming more and more difficult. The use of sensors to detect process faults and to allow for process control are increasingly viewed by semiconductor manufacturers as a means to improve yields more rapidly and also as a means to significantly improve equipment productivity. New sensors and sensor application development are needed but this is only part of the problem. Process tool controllers that can easily incorporate new sensors into manufacturing tools are needed NOW. To that end SEMATECH has initiated several projects to development new sensors, develop sensor applications, and accelerate the availability of new process tool controllers that will allow for easy integration of new sensors and control algorithms. An overview of these activities will be presented.
2:10 PM		MS-MoA-3 Fault Detection and Process Control under Development in the University Community - What is Possible in the Future? C. Spanos (University of California, Berkeley) This presentation will outline the work of the Berkeley Computer-Aided Manufacturing group on real-time plasma monitoring and control, as well as in run-to-run lithography control and diagnosis. These projects involve a collection of the best available sensors (RF, OES, particle counters, in-situ reflectometry, etc.) along with novel data processing algorithms, such as real-time statistical process control, adaptive supervisory control, etc. This work has lead to the capability to forecast machine failures, and, in this way, extent the mean time to failure figure. We have also demonstrated the ability to correctly diagnose problems from in-situ and also from real-time signals, and in this way we have reduced the mean time to repair figure. Current work is geared towards run-to-run and real-time control that is expected to enhance the processing capability (Cpk) for deep submicron patterning.The body of knowledge gained during these last ten years is now pointing to the future challenges, as semiconductor production technologies move to a tenth of a micron features and larger, more expensive wafers. Currently, the most critical issues involve modeling robustness and sensor availability and performance. Several promising sensor techniques must be examined and judged on their capability to support automated control and fault detection applications. Further, the fundamental error budget of 0.1\mu\m technologies must be analyzed in order to determine the specifications of future sensing and modeling techniques, as these techniques will be driving the controllers of the future semiconductor factory.
2:50 PM		MS-MoA-5 Experimental Study of RGA High Pressure Performance (II) M. Li, S. Tison, P. Abbott, C. Tilford (National Institute of Standards & Technology) In the talk we presented at last year's AVS national symposium, we reported that the RGA sensitivity strongly depends on the gas pressure, particularly at the high pressures (above 10\super -3\ Pa) where RGAs are often used to monitor and control semiconductor processes. We believe this is an ion space charge effect, and we have experimentally demonstrated that instrument parameters that affect space charge will affect the pressure dependence of the RGA sensitivity. In particular, the sensitivity decreases as the electron emission current decreases and/or the ion energy setting increases. For one instrument at an ion energy setting of 25 eV, the sensitivity variation in a pressure range between 10\super -6\ to 10\super -1\ Pa is reduced to a factor less than 3 at an emission current of 0.01 mA compared to a factor of 33 at 2 mA. Other instruments tested show a similar improvement in the sensitivity linearity when operated at low emission currents. It has also been observed that increasing the ion energy setting affects the sensitivity nonlinearity significantly when the emission current is larger than 0.5 mA. Our efforts in modeling of the effects of ion space charge outside the ion source on the RGA sensitivity continue. We have examined the perturbation of the electric quadrupole field near the quadrupole filter entrance due to the ion space charge, and this has shown that the amplitude of the stable ion trajectories in an ideal quadrupole field is increased because of this ion space charge. The resulting decrease in the RGA sensitivity were calculated numerically and compared with the results of sensitivity measurements.
3:10 PM		MS-MoA-6 Equipment and Process Fault Detection and Classification in PolySi RTCVD L. Tedder, G. Lu, N. Rabbani, G. Rubloff (North Carolina State University) Active sampling mass spectrometry has been previously applied to capture, in real-time, the time-dependent behavior of equipment, process, and wafer state conditions during the entire RTCVD process cycle. In this work, we have experimentally induced various faults and observed their consequences in the mass spec data, including: a sticky and broken mass flow controller (MFC), an incorrect MFC setting, the freezing of throttle valve control for total pressure, and the mis-calibration of regulatory control parameters for total pressure and wafer temperature. The mass spec data display evidence of these faults clearly in real time. Other equipment parameters (e.g., throttle valve position, pyrometer reading, process timing) add information crucial for fault classification, especially for distinguishing system faults >from sensor faults. Dynamic simulation, which represents time-dependent equipment, process, sensor, control system, and wafer state information, is used for comparison to experiment in an effort to develop decision-making methodologies for systematic fault classification.
3:30 PM		MS-MoA-7 Stabilization of a Plasma-enhanced CVD Process using Quadrupole Mass Spectrometry T. Knight, X. Cheng, D. Greve, B. Krogh (Carnegie Mellon University) We have implemented Quadrupole Mass Spectrometry sensing in a parallel-plate plasma-enhanced CVD silicon nitride process to measure gas species concentrations in real time. We have shown in recent work that the gas species concentrations chosen (disilane and triaminosilane) are closely related to film properties. A Linear-Quadratic Gaussian (LQG) controller has been designed based on black-box system identification experiments in order to drive system inputs such as gas flows and chamber pressure. The controller manipulates gas species concentrations and hence film properties. This makes the system tolerant to common disturbances. An experiment was performed in which the silane mass flow controller was intentionally miscalibrated by adjusting it to various zero offset levels simulating a reasonable system disturbance. Films deposited without the real-time controller displayed index of refraction and deposition rates which varied as a function of the zero offset value. However, in depositions with the same zero offset values but using the controller to provide feedback, the controller compensated for the disturbance and achieved consistent film properties regardless of the mass flow controller offset. By using in situ sensing and control, the process can defined in terms of process variables which are more closely related to film properties than inputs such as mass flow rates and RF power.
3:50 PM		MS-MoA-8 Process Control and Diagnostics by Means of Self Excited Electron Resonance Spectroscopy (SEERS) M. Klick, M. Miscke, W. Rehak, D. Suchland (Adolf-Slaby-Institute, Germany) Nonlinear phenomena in asysmmetrical rf discharges are known for many years. In modeling rf discharges nonlinear phenomena usually are treated as inconvenient effects, e.g. concerning the application of linear models. It will be shown without a kinetic approach that nonlinear mechanisms play a dominant role. The nonlinearity of the sheath at the powered electrode and oscillations in the plasma are analysed for the RF discharge. For asymmetrical discharges and sinusiodal discharge the current can be shown to consist of a saw tooth shaped part and a superposed damped oscillation. This model results in a new method for plasma monitoring and allows to estimate density and collision rate of the electrons in the plasma body. It is called self excited electron resonance spectroscopy (SEERS) and can be used for rf plasma monitoring and process control. Microwave interferometry and Langmuir probe maesurements were used to verify this method. A CMOS-gate etching process using RIE is utilized to show how plasma parameters (e.g. electron density) depend on the chemical surface reaction. The high sensitivity and the capability of endpoint detection are shown.
4:10 PM		MS-MoA-9 Run by Run Uniformity Control on a Dual Coil Transformer Coupled Plasma Reactor with Full Wafer Interferometry M. Le, D. Boning, H. Sawin (Massachusetts Institute of Technology) Etch rate uniformity can be controlled in a dual coil transformer coupled plasma (TCP) with Full Wafer Interferometry (FWI) as a process monitor. Single coil TCPs offer many of the advantages of high density low pressure plasmas but are subject to uniformity limitations. This variation is primarily due to nonuniform RF power coupling but is exacerbated because the etch rate and uniformity tend to drift with processing. Present commercial TCP reactors utilize a single optimized inductive coil of fixed geometry. Computer simulations (Paranjpe) have shown that a single flat coil of small radius tends to produce nonuniform etch profiles peaked in the center while a coil of large radius will result in off axis peaked etch profiles. As suggested by Yamada et al., by splitting a single coil into two separate coils, one of small radius and one of large, the etching rate uniformity can be tailored by varying the fraction of total RF power delivered to each of the concentric coils. As the process drifts, corrections can be made to the process settings to compensate for deviations in the etching uniformity. Two separate coils with independent power supplies and matching networks have been installed on a Lam TCP. Blanket polysilicon films on 6 inch wafers were etched with Cl/sub 2/2 and HBr gases. Full Wafer Interferometry (FWI) is used to measure the etching rate and uniformity by observing the modulation of the plasma induced emission due to interference as the film is etched. This in situ diagnostic can be coupled to a run by run controller to adjust the split coil process conditions for the next wafer as to optimize etching uniformity. Our results demonstrate the applicability of using a dual coil design in conjunction with measurements made with FWI for improved uniformity control.
4:30 PM		MS-MoA-10 An Artificial Neural Network EWMA Controller for Semiconductor Processes T. Smith, D. Boning (Massachusetts Institute of Technology) The exponentially weighted moving average (EWMA) controller has previously been shown to improve semiconductor process control for approximately linear processes which are subject to shifts or persistent drifts in the presence of noise. This work addresses the inability of the basic EWMA controller to adequately control processes which are poorly modeled by a linear approximation. This issue is imperative to the success of the EWMA controller in semiconductor manufacturing where many processes in reality exhibit substantial nonlinear behavior. We address this issue by extending the EWMA controller to utilize and adapt an artificial neural network (ANN) process model. The ANN model is dynamically updated using an EWMA of the biases in the ANN output layer. Recipe generation takes place by outimizing around the dynamic ANN model. We demonstrate that this framework greatly improves the performance compared to an EWMA controller for higher order processes. Simulation results indicate that this controller provides stable control for higher order processes which would cause the affine EWMA controller to go unstable and fail to control the process. In addition, we suggest that this ANN/EWMA architecture is very robust in the face of model error and noise. The key contribution is the integration of two key process control capabilities. First, the ANN/EWMA system retains the benefit of an EWMA controller, namely good control with minimal added process noise. Second, this ability to control in the face of noise is retained for higher order complex nonlinear semiconductor processes.