Optimizing STI etch process control using optical digital profilometry

Robert Chong and Kevin Lensing, AMD; Thomas F. Edgar, University of Texas; and Bryan Swain, Timbre Technologies

Over the last three years, FASL, a joint venture between Fujitsu (Tokyo) and Advanced Micro Devices (AMD; Sunnyvale, CA), has pursued an aggressive technology roadmap at AMD's Fab25 in Austin, TX. In 2002, the fab, which manufactures Spansion flash memory chips, began to convert from microprocessor to flash memory production. In 2003, it successfully completed both a steep production ramp and a technology shrink from 170 to 130 nm. Spansion's floating-gate flash memory has ramped into high-volume production at 110 nm. Short-term challenges include the qualification of the 90-nm technology node. As the size of the memory cell decreases and the packing density increases, it is becoming more and more vital to ensure that each bit is electrically isolated from its neighbors. Failure to electrically isolate the device core can make it difficult to program and erase the memory cell or, even worse, create reliability issues as a result of static-charge loss.

Figure 1: Schematic diagram of a postetch STI trench.

The STI Process

In VLSI semiconductor manufacturing, shallow-trench isolation (STI) is the most commonly used method for separating electrically active device regions. The STI process involves patterning and etching shallow trenches into the silicon substrate of the wafer. During the masking step, STI resist is patterned on top of blanket layers of silicon oxynitride (SiON), silicon nitride, and silicon dioxide. The wafers are then processed in a plasma etcher to completely etch away the open regions of the SiON, nitride, and oxide, after which a targeted etch is performed on the underlying silicon. The resulting trenches, depicted in the schematic diagram in Figure 1, are then filled with dielectric material to form isolation barriers between adjacent electrical structures. The STI process is considered to be a critical manufacturing step because it constrains all subsequent patterning layers by defining the active regions for the electrical structures.

Since the bulk silicon has no underlying stop layer, normal variations in the silicon etch rate of the process tool during STI processing lead directly to variations in the silicon trench depth. To minimize these variations, Fab25 implemented an advanced process control (APC) run-to-run (R2R) controller in the silicon etch portion of the STI process. AMD and FASL have pioneered the deployment of APC as a part of their automated precision manufacturing (APM) program, which has enabled the rapid technology program used at Fab25. R2R controllers such as that used to perform STI trench etch at Fab25 help the maturation of new technologies by effectively eliminating systemic variation from the lot-mean controlled process variable.

The STI trench-etch R2R controller is based on the model-inverse controller design. However, the APM group discovered that the stability of the process can be improved significantly if the metrology data stream is carefully managed by means of advanced filtering, proper tool selection, and factory-level integration between different manufacturing systems. At AMD/FASL, the concept of data management has been extended from databases and brokers to a conglomeration of high-level systems that are directed at maximizing the utilization of data streams. This article demonstrates that the metrology data stream from an optical digital profilometry (ODP) system supplied by Timbre Technologies (Santa Clara, CA) can significantly improve the performance of the fab's STI etch R2R controller.

Figure 2: Schematic diagram of the trench-etch controller.

Trench-Etch R2R Control Algorithm

Figure 2 presents a schematic diagram of the trench-etch R2R controller, which is based on a model-inverse controller and has an exponentially weighted moving average (EWMA) state estimator. The state estimator is used to forecast the next state, which is then used to calculate the next control inputs to the process tool. The form of the state estimator is shown in the following equation:

where forecasted state for time t; λ = the tuning parameter; z_t = measured state obtained at time t; and = forecasted state at time t – 1).

For the trench-etch process, the EWMA state estimator has many attractive properties, making it suitable for use in several semiconductor processes. First, it enables compact forecast calculations and contains only one tuning parameter to account for the controller's performance. Second, it provides a forecast with a minimum mean square error, which allows the long-run average of the squares of the forecast errors (e_t) to converge to the smallest possible value. Thus, for a process with white noise (a_t) and a zero mean, the mean (long-run average) of the e_t values is equivalent to zero because e_t converges to a_t. Hence,

This equation also implies that the forecast errors will have the smallest possible variance (i.e., that variance, σ^{2 _e} , will be minimized. Hence, the standard deviation, σ^{2 _e} , will also be minimized.

While these desirable properties make the EWMA state estimator a very attractive tool for use in control applications, the state estimator's performance can be impaired if sufficient plant-model mismatches exist. Plant-model mismatches typically occur when the mathematical models insufficiently account for the dynamic behavior of the physical system (i.e., the plant). In Fab25's trench-etch R2R controller, the plant-model mismatch was pronounced enough within the model conversion that the robustness properties of the EWMA state estimator could not compensate for it. As a result, the trench-etch process was suboptimal. The plant-model mismatch was caused by an inaccurate estimated parameter used in the process model. The mismatch affected STI process controllability.

Developing an Improved Process Model

Original Metrology Model. The original process model employed by the R2R trench-etch controller targeted the depth of the silicon trench based on feedforward thickness measurements of successive film layers:

Original_Model = f (TrenchDepth,FeedForward₁ . . . FeedForward_n,OverEtch,Time)

The feedforward variables of the original process model were driven by metrology requirements. In order to estimate the silicon etch depth, the controller subtracted the thickness values of previously deposited thin films from the total trench depth as measured by a surface profilometer. In theory, that method is reasonable for estimating the amount of silicon that was removed. But in reality, it has several drawbacks.

Figure 3: Schematic diagram of the movement of a profilometer stylus over an STI trench.

The first is that profilometer technology is based on measuring the deflection of a stylus affixed to the tip of a cantilever beam. As the stylus is physically dragged across the surface of the wafer, as illustrated in Figure 3, it records the step height of the structures it traces. This method is not robust enough to perform measurements on modern technology nodes. At Fab25, profilometer gauge studies have shown that the instrument's precision is on the order of only 2.0 nm. Since device specifications for new technologies allow for a total STI trench variation of ~10 nm, the profilometer's noise is unacceptable.

In addition to poor repeatability, the profilometer is blind to process disturbances in prior thin-film layers. In subtracting the previously deposited thin-film measurements from the total step height to estimate the amount of silicon etched, the process model assumed, first, that the feedforward metrology measurements of prior film layers were always available, and, second, that both the process and metrology noise in the feedforward measurements were centered around zero. However, these assumptions were not always valid. The measurements were subject to a limited metrology sampling plan, and feedforward metrology was sometimes biased by either process or metrology noise from the film layers. Those errors introduced a cross-correlation effect into the process model, signifying that unmodeled dynamics were still present.

Table I: Correlation coefficients for the original process model.

Feedforward metrology disturbances can be observed by evaluating correlation coefficients between measured trench depth and feedforward films. Table I presents correlation coefficients for the original process model, illustrating that trench depth was influenced by incoming film measurements. For example, trench depth had a negative correlation to feedforward film 1 but a positive correlation to feedforward film 2. An optimal process model would be sensitive to disturbances from both film layers. But to achieve the required sensitivity with the original model, the model would have to have been unduly complex.

Figure 4: Cross-correlation between the calculated etch-time and the measured trench depth from the original process model.

To determine how complex such a model might have been, a time-series analysis for the original process model was performed using the cross-correlation between calculated etch time and the measured trench depth. The results of that analysis, presented in Figure 4, indicate that there was significant coupling between output and input. Furthermore, process gains, corresponding to the prominent peaks, appeared to fluctuate erratically between positive and negative. Finally, the strongest output effect appeared to be at a sample interval of approximately 23, indicating that an overly complicated process model was necessary to capture the process dynamics introduced by the feedforward film measurements.

Simplified Process Model. To enhance the capabilities of the metrology process, the ODP metrology system was evaluated. ODP is a novel technique based on the science of scatterometry. Utilizing an optical signal from a spectroscopic ellipsometer or reflectometer, it delivers a digitized cross-sectional profile of semiconductor device features. ODP profiles contain the geometric details of the patterned structure, including shape, critical dimension (CD), and film thicknesses. The system works by matching optical spectra from a periodically repeating test (grating) structure on a process wafer with a theoretically modeled solution of Maxwell's equations based on the 0th order diffraction pattern of the optical signal.

At Fab25, ODP measurements were performed using an Optiprobe 5230 spectroscopic ellipsometer from ThermaWave (Fremont, CA). The system delivers both intensity (tan ψ) and phase (cos Δ) information over a continuous wavelength range from 193 to 1100 nm. A pictorial overview of the system is presented in Figure 5.

Figure 5: Schematic diagram of ODP technology flow.

To perform ODP-based metrology, a reference library of theoretical solutions that encompasses the process window of the patterned feature under consideration must be generated. To create that library file, the user inputs measured spectra and expected structure characteristics such as pitch, typical process ranges, and nominal CD and film-thickness values into the ODP software. Based on those inputs, ODP-NOW software regressively determines a solution set for the manufacturing process and suggests a parameter set to define the process space. Once the file has been created using library-generation software, it is installed on a profile applications server (PAS) and associated with an appropriate recipe corresponding to the device of interest.

In a production environment, wafers are measured using the spectroscopic ellipsometer, and the generated spectra are delivered to the PAS. The sample data are then compared with the solution set in the library file to determine the optimized match. Once that match has been determined, the PAS outputs the corresponding digital profile and its geometric attributes.

Figure 6: ODP profile of an etched STI trench feature superimposed over a corresponding SEM image.

The ODP system outperforms the profilometer metrology system. Fab25's gauge studies determined that ODP's precision is on the order of 0.67 nm, which is three times greater than that of the profilometer. In addition, the system offers advanced modeling and filtering capabilities that are not available on the profilometer. Its modeled outputs are latent images of trench profiles that accurately depict the shape and dimensions of an STI trench. Figure 6 illustrates an ODP profile of an etched STI trench feature superimposed over a corresponding cross-section scanning electron microscope (SEM) image. The ODP system can also model the patterned film layers on top of the silicon substrate, allowing the trench-etch R2R controller to calculate only the amount of silicon removed from the STI trench film stack. The system thereby reduces controller variability by ignoring incoming disturbances associated with feedforward metrology measurements.

The ODP system not only segregates individual films within a composite film stack, it also provides a confidence value for the metrology output in the form of a goodness of fit between the sample spectra and the matched solution. In addition, it generates an alarm when the sample data fall outside the process range as specified by the reference library. Thus, in contrast to the original metrology system, the R2R controller can filter out metrology errors and process faults.

By combining accurate models with advanced filtering techniques, the ODP system made it possible to simplify the original process model. The new model,

Simplified_Model = f (TrenchDepth, OverEtch,Time)

enables greater process flexibility by eliminating feedforward metrology dependencies, which in turn helps to stabilize the process.

Process Results

Rather than develop overly complicated process models to compensate for the peaks shown in Figure 4, ODP was implemented to ignore them altogether. Since the feedforward film measurements were no longer used in the new model, process dynamics should reflect that change. That hypothesis can be investigated by evaluating the cross-correlation results from the simplified process model.

Figure 7: Cross-correlation between the calculated etch-time and the measured trench depth from the simplified process model.

Figure 7 indicates that process gains from the new model were stabilized, since they no longer fluctuated erratically between positive and negative, as in the original model. In addition, the strongest effect on the output series appeared to be at a sample interval of approximately 16. Hence, the new model required seven terms fewer than the original one would have required, demonstrating that the new model is less complex.

The most important benefit of implementing the ODP system was that it tightened the silicon etch-rate standard deviations by approximately 38%—a substantial improvement for any manufacturing process. Figure 8 shows normalized etch-rate histograms for the original versus the simplified process model.

Figure 8: Etch-rate histograms from the original and simplified-process models.

In addition, the ODP system was able to improve metrology tool capacity in two ways. First, it reduced upstream metrology requirements associated with the feedforward elements in the R2R process model. Metrology sampling at Fab25 for STI blanket films is now based on the process control needs of the deposition processes alone rather than on the downstream requirements of trench-etch APC. Second, the ODP system is approximately 40% faster at measuring three wafers with nine sites each than the profilometer. The increased speed enables the controller to receive metrology updates more quickly, increasing wafer throughput and ultimately improving cycle times. These improvements ease fab engineering requirements, increase metrology efficiency, and reduce latency effects to achieve better process control.

Conclusion

ODP is a novel method for performing nondestructive dense-array STI trench metrology. The system can perform high-definition inspection of individual films within a complex stack without sacrificing precision or throughput. By using ODP software to model the signal from a spectroscopic ellipsometer, the Fab25 STI trench-etch R2R controller was optimized for maximum performance. The ODP-based controller stabilized the process substantially. With superior signal-to-noise levels, improved throughput, and more-robust output metrics, the system is an excellent data stream for R2R control solutions in high-volume manufacturing facilities.

Acknowledgments

The authors would like to thank their managers at AMD, FASL, and Tokyo Electron America (TEL) in Austin, TX, for providing the time and resources to support projects of this nature. In addition, the authors would like to thank Michelle Pesez of TEL for her work in making this article come to fruition.

Bibliography

Bode, C, and T Sonderman. "Controlling the Margins in 300 mm Manufacturing." Solid State Technology 47, no. 2 (2004): 49–52.

Box, G, G Jenkins, and G Reinsel. Time Series Analysis: Forecasting and Control, 3rd ed. Upper Saddle River, NJ: Prentice-Hall, 1994.

Box, G, and A Luceño. Statistical Control by Monitoring and Feedback Adjustment. New York: Wiley, 1997.

Lensing, KR. "Scatterometry Feasibility Studies for 0.13-Micron Flash Memory Lithography Applications; Enabling Integrated Metrology." Paper presented at SPIE Microlithography 2004, San Jose, February 22–27, 2004.

Lensing, KR, et al. "Shallow Trench Isolation Scatterometry Metrology in a High Volume Fab." In Proceedings of the ISSM 2001. Piscataway, NJ: IEEE, 2001, 195–198.

Seborg, D, T Edgar, and D Mellichamp, Process Dynamics and Control, 2nd ed. New York: Wiley, 2004.

Robert Chong is a senior process development engineer in the automated precision manufacturing (APM) organization at AMD in Austin, TX. Before joining AMD, he was a process engineer for metallization at Atmel in Colorado Springs, CO. He received a BS in chemical engineering from the University of Texas (Austin) and is pursuing a master's degree in process control. (Chong can be reached at 512/602-1308 or robert.chong@amd.com.)