Defect inspection technologies to meet the challenges of advanced CMP process
03/01/1998
Defect inspection technologies to meet the challenges of advanced CMP processes
James Reynolds, Reynolds Consulting, Sunnyvale, California
Aaron L. Swecker, Andrzej J. Strojwas, Carnegie Mellon University, Pittsburgh, Pennsylvania
Ady Levy, Bobby Bell, KLA-Tencor, San Jose, California
Fast, accurate inspection for defects caused by CMP processes is essential to an effective yield management program. Dark-field system parameters such as illumination angle, polarization, and the effects of repetitive pattern filtering (Fourier filtering) are studied using simulation and experimental techniques. For grain noise reduction, a low angle of illumination produced results superior to a high angle of illumination. A repetitive pattern filter effectively improved detection in the presence of pattern noise. With color noise present, circular polarization of the incident illumination provided a higher signal-to-noise ratio than p or s polarization.
Integrated circuit processing continues to become more complex as new processes and more steps are required to fabricate advanced devices. As a result, in-line yield management using optical automated defect inspection is becoming increasingly important for faster-yield ramps and higher-yield stability.
Recent advances in chemical mechanical polishing (CMP) have allowed additional metallization layers and tighter packing densities through improved planarization. Three types of CMP processes are widely used: shallow trench isolation, oxide CMP, and metal CMP.
With the advent of CMP, in-line yield management using automated inspection has faced new challenges, including new yield-limiting defect types (e.g., microscratches) and additional process noise sources. Noise sources at post-CMP inspection now include color from oxide thickness variation, underlying and top surface pattern variations, and grain noise from metal and polysilicon.
In order to maximize sensitivity and capture rate of critical defects, and to improve the robustness of the in-line inspection recipe, an important design objective is to maximize signal from the defect, while consistently suppressing the noise. This drives careful study and optimization of the architectural design of inspection systems.
 |
Figure 1. Color noise caused by CMP oxide thickness variation on four adjacent die [1].
This article studies inspection tool parameters, such as dark- field illumination angle, polarization, and spatial filtering, to find ways of improving CMP defect capture rate in the presence of these noise sources. We investigated two critical CMP defects, the microscratch and residual slurry particle, and used rigorous 2-D and 3-D electromagnetic (optical) simulations to ascertain the optimal configuration for each inspection parameter. In addition, experiments are used to validate the simulation results.
CMP inspection challenges
Successful yield management in CMP requires detection of all critical defects in the presence of the high noise sources from advanced processing. Widely seen critical CMP defects can be separated into two categories: foreign material such as residual slurry, surface particles, embedded particles, and residual tungsten; and voids in the surface material such as microscratches, ripouts, and dishing.
The three prevalent sources of noise that impact defect inspection at post-CMP processing are layer thickness variation (color noise), grain noise, and pattern noise. Silicon dioxide (SiO2) layer thickness variation (Fig. 1) is introduced by nonuniform polishing across the wafer surface. Depending on inspection tool architecture, this variation can cause fluctuations in the magnitude of light scattering from the topography and defects, making it difficult for the system to separate the signal of the defect from the noise reliably.
 |
Figure 2. Grain noise on underlying metal causes dark spots that make defect detection more difficult [2].
Grain noise from underlying and surface metals, such as aluminum, tungsten, and polysilicon, is becoming more significant as grain sizes increase with advanced processing. This noise results from randomly distributed grains that are not completely subtracted when two adjacent die or cells are compared. Figure 2 shows a representative image of this grain noise. Depending on the inspection tool design, these random dark/bright grain spots (noise) can interfere with the detection of critical defects.
Pattern noise is present when the surface pattern on the wafer interferes with defect detection. This can result from pattern variation, such as linewidth variation, or strong scattering from the pattern that limits the available detection dynamic range. The inspection tool should suppress the nondefective pattern and maximize critical defect capture.
Inspection system architectures
Two common architectures are used for in-line automated defect inspection: bright field and dark field. Bright-field illumination systems collect the reflected and scattered light from the pattern to resolve the defects, taking advantage of contrast differences between the defective and nondefective regions. Since reflected light is used in bright-field inspection, one of the main challenges for CMP inspection has been the reduction of interference between the oxide surface reflection and the underlying layer reflectance. This interference strongly depends on the thickness of the oxide layer that varies with CMP processing, resulting in large signal differences across the wafer, appearing optically as color variation or noise.
An ultra broadband illumination source dramatically reduces the amplitude of these color variations [1]. For example, Fig. 3 shows the difference in signal intensity as a function of oxide layer thickness, comparing the ultra broadband source with a monochromatic source of illumination. For all collection magnifications (numerical apertures), ultra broadband is superior to monochromatic bright-field illumination in suppressing color noise [3].
In contrast to bright field, dark-field systems collect only the scattered light, while rejecting the spectrally reflected light. A dark-field system consists of an illumination source (typically a laser) at an incidence angle q, an input polarizer, and optics that collect the scattered light. The input polarizer selects either s (parallel to the wafer plane), p (parallel to the illumination plane), or circular polarization. Some systems include a repetitive pattern filter in the collection optics path for pattern filtering.
We used the simulation package METRO [4, 5] to study the effect of these dark-field system attributes on reduction of CMP noise sources and improvement of defect capture. METRO is a physically based electromagnetic field solver that calculates the reflected and transmitted light from wafer topography. It is a rigorous vector diffraction method that solves Maxwell`s equations and, therefore, can model in-line inspection schemes of interest. We then used an experimental test bench to validate the simulation results.
 |
Figure 3. a) Bright-field monochromatic and b) ultra broadband signal of SiO2 over the Si stack as a function of oxide layer thickness for various numerical apertures. Black, NA = 0.25; red, NA = 0.45; green, NA = 0.75; and blue, NA = 0.90.
Simulation setup
Two defects are studied in this paper: microscratches and slurry particles. Microscratches can occur during CMP when a particle is dragged across the wafer surface, causing a shallow void in the oxide layer. Slurry particles are the abrasives used during CMP.
To simulate the grain noise, we randomly distributed metal pyramid-shaped scattering centers with 0.3-?m bases and heights on top of a metal film. The residual slurry particle was modeled as a 0.6-?m polystyrene latex (PSL) sphere. For the evaluation of the effect of color pattern noise on defect detection, we modeled the microscratches as trenches in the oxide with a depth (d) of 25 nm and a width (W) of 0.5 ?m, and the slurry particle as a PSL sphere with a diameter of 0.1 ?m. We simulated the microscratch and the particle on a SiO2 layer of various thicknesses (h) over both a polysilicon-SiO2 grating or pattern on a silicon substrate, and a uniform silicon layer or film. The pitch was 1 ?m with 0.35-?m polysilicon lines and 0.65-?m SiO2 spacing for the polysilicon grating.
Results and analysis
Grain noise reduction. We evaluated the detection of residual slurry particles in the presence of grain noise for low (grazing or near the wafer surface) and high angles of illumination (approximately 45? to the wafer surface). Figures 4a and 4b display the normalized intensity for the high and low angles of illumination. Using a low angle of illumination, the defect signal is substantially higher than the surrounding grain noise. However, at a high angle of illumination, the residual slurry particle signal is lower than the grain noise, and therefore undetectable.
The enhanced detection of the particle using the low angle of illumination is attributed to the creation of a standing wave near the wafer surface. There is a null in the incident intensity near the surface and a peak at a height that is inversely proportional to the sine of the illumination angle. At a low angle of illumination, the intensity peak (and therefore the scattering intensity) is located above the grain, resulting in weak grain noise. In contrast, at a high angle, the incident light intensity peak of the defect is located at the grain height, which enhances the grain scattering and reduces the signal-to-noise ratio of the defect (Fig. 5).
 |
Figure 4. Simulation results for a particle on a grainy surface at a) high and b) low angles of illumination. In the low-angle case, the defect signal is clearly distinguished from the grain noise.
 |
Figure 5. Experimental results for a particle on a grainy surface at a) high and b) low angles of illumination. In the low-angle case, the particle is clearly visible above the grain noise.
 |
Figure 6. Simulation results for microscratch and particle defects a) with and b) without the repetitive pattern filter. Defect signals are easily distinguished from the pattern when the filter is present.
Pattern noise reduction. The dynamic range of the system limits the detection of small defects in the presence of strong background scattering, such as that from pattern. Optical removal of the pattern signal prior to collection by the sensor can therefore enhance the sensitivity of the detection tool. For repetitive patterns (such as a DRAM or cache array in a microprocessor), we can place a mask (Fourier filter) in the collection optical path.
The scattered light from a repetitive structure only propagates in discrete angles (orders) [6]:
sin(qn) - sin(q0) = nl/P
where q is the angle measured from the normal of the wafer, n is the integer order number, l is the wavelength of light, and P is the period of the structure. Order number 0 corresponds to specular reflection. Blocking these orders using a mask can eliminate the pattern signal, while still collecting the nonrepetitive defect signal.
Figure 6a shows the scattered intensity from a dense, repetitive-patterned array and two defects: a microscratch and a slurry particle. The pattern signal is much larger than the signal from the two defects, therefore limiting the available dynamic range for detection of those defects. Figure 6b presents the normalized signal from the microscratch and the particle after removing the repetitive signal optically using a repetitive pattern filter. In this case, the dynamic range of the system can be tuned to detect much smaller defects, as demonstrated by the improved signal-to-noise ratio of a microscratch in an oxide CMP film over an advanced DRAM array in Fig. 7.
Color noise reduction. We simulated the effects of layer thickness variation on dark-field detection using s, p, and circular polarization at an oblique angle of illumination. The results show strong dependence of the scattered light from the defect on both the incidence light polarization and the oxide layer thickness.
The objective of an effective detection system is to detect the critical defects regardless of layer thickness. Figure 8 plots the noise limit or minimum detection limit for each polarization. For both microscratch and slurry particle defects, we obtained the highest noise limit or tolerance level using circular polarization [7].
 |
Figure 7. Experimental results showing signal-to-noise enhancement for microscratch over pattern a) without and b) with a repetitive pattern filter.
 |
Figure 8. Simulation results of dark-field detection using s, p, and circular polarization on a) a slurry particle and b) a microscratch. We obtained the highest noise limit or tolerance level using circular polarization [7].
 |
Conclusion
We studied grain, pattern, and color noise reduction techniques for enhanced, dark-field defect detection on advanced CMP processes. The table summarizes the dark-field system parameters and optimal conditions for inspection. Using simulations, we have shown that dark-field low angle of illumination enhances the detection of slurry particles in the presence of grainy CMP layers; optical repetitive pattern filtering increases the detection sensitivity in dense repetitive arrays; and circular polarization increases the noise tolerance limit in the presence of color noise.
References
1. S. Zika, KLA Yield Management Seminar, San Jose, CA, March 1997.
2. D. Schmidt, KLA Yield Management Seminar, San Jose, CA, March 1997.
3. A. Swecker et al., "Comparison of Defect Detection Schemes Using Rigorous 3-D EM Field Simulations," Proc. of SPIE Metrology, Inspection, and Process Control for Microlithography XI, Vol. 3050, pp. 313-321, March 1997.
4. C. Yuan, "Efficient Light-Scattering Modeling for Alignment, Metrology, and Resist Exposure in Photolithography," IEEE Trans. on Electron Devices, Vol. 39, No. 7, pp. 1588-1598, July 1992.
5. K. Lucas, A. Strojwas, H. Tanabe, "Efficient and Rigorous 3-D Model for Optical Lithography Simulation," J. Opt. Soc. Am. A, Vol. 13, No. 11, pp. 2187-2199, Nov. 1996.
6. M. Born, D. Wolfe, Principles of Optics, pp. 401-405, Pergamon Press, NY, 1993.
7. A. Swecker, A. Strojwas, A. Levy, B. Bell, "Evaluation of Defect Detection Schemes for CMP Process Monitoring Using Rigorous 3-D EM Simulations," Proc. of ASMC, pp. 283-288, Sept. 1997.
JAMES REYNOLDS received his BA degree in physics from Williams College, and his MBA degree from The University of Santa Clara. He is president of Reynolds Consulting, a firm founded in 1981 to offer technical consulting services to the lithography community. Reynolds Consulting, 1110 Sunnyvale-Saratoga Rd., Sunnyvale, CA 94087; ph 408/732-6275, fax 408/732-6370.
AARON L. SWECKER received BS degrees in electrical and computer engineering at West Virginia University in 1995, and his MS degree in electrical and computer engineering at Carnegie Mellon University (CMU) in December 1996. He is pursuing his PhD degree at CMU under the direction of Andrzej Strojwas.
ANDRZEJ J. STROJWAS received his PhD degree from CMU in 1982. He is a professor in the Electrical and Computer Engineering Department at CMU and a fellow of IEEE. Strojwas has published more than 300 papers on semiconductor manufacturing and VLSI CAD.
ADY LEVY received his PhD degree in solid state physics from the Massachusetts Institute of Technology in 1992, and worked as a postdoctoral researcher at IBM T.J. Watson Research Center until 1995. He is the optical system manager of the wafer inspection business unit at KLA-Tencor.
BOBBY BELL received his MSEE degree from the University of Missouri in 1986, and his BSEE degree from the University of Arkansas in 1984. Bell is the director of marketing for the wafer inspection business unit at KLA-Tencor.