See-through-silicon Inspection Application Studies Based on Traditional Silicon Imager Wei Zhou, Darcy Hart, Noah Bock and Rolf Shervey - Rudolph Technologies Presented at SPIE Optics & Photonics, August 2011 With semiconductor development processes hitting harder and harder on Moore’s law to continuously scale down, high- density advanced packaging technologies become a promising alternate route to improve transistor density. Chip integration IO/cm2 density jumps quickly by orders from 2D packaging of 102 to wire bonded chip stack of 103, to TSV of 104~105 and to advanced 3D integration of 105 to 106. Starting with wire bonding and now prevailing with TSV, more and more silicon layers are stacked up in 3D dimension to improve system density. A typical stacked wafer sample has two wafers glued together with a patterned area sandwiched in between. Outer surfaces can be polished or unpolished bare silicon surface or patterned surface. See-through-silicon inspection is becoming an essential quality control method for semiconductor manufacturing in three-dimensional integrated circuit (3DIC) and chip-stacking and related industries such as crack inspection in solar panel manufacturing. Due to silicon’s band-gap structure, silicon wafers are transparent to light beyond 1.2 micron wavelength range, so silicon-based detectors have very low sensitivity to this wavelength range. InGaAs detectors are typically favored for such applications, even with much higher cost, larger pixel size and smaller detector resolution. This paper discusses a series of studies to improve silicon-based imagers’ detection signal-to-noise ratio for this see-through-silicon inspection application. Silicon-based detector sensitivity increases when light wavelength shifts shorter from 1.2 micron. Both photon penetration depth and percentage of penetrated photons, however, drop accordingly. Non-throughpenetrated photons are scattered in all directions both at the wafer top surface and inside the silicon material. A portion of scattered photons will be collected by inspection optics and reach the imager eventually, causing a white noise background in images out of silicon based detectors, as shown in figure 1. This essentially results in low image contrast. Figure 1 - Raw 8bit BMP image from CMOS imager Figure 2 - Processed image (Wafer courtesy of SEMATECH) Effectively removing this background noise is the key to fundamentally improve the signal-to-noise ratio of silicon-based detectors for this application. A series of methods were studied, including simple image processing methods, advanced frequency domain analysis and reference subtraction, and each method’s pros and cons were compared for industrial inspection applications. Figure 2 shows an example of processed images in this study, which carries better SNR and visibility for both machine vision inspection and human vision review. SEE-THROUGH-SILICON FUNDAMENTALS As shown in figure 3, silicon material’s band gap structure results in a very high absorption rate in visible wavelength range; very minimum photons can transmit through silicon material in the wavelength range from UV light to NIR region (0.3~1.0um). Only low-energy photons with wavelengths larger than 1.2um can penetrate through silicon material to reach the other side of a silicon wafer, and a portion of photons will be reflected back though silicon material again. Therefore, InGaAs detectors are naturally the best candidate for this see-through-silicon application, as shown in figure 4. Unfortunately, InGaAs detectors have a much larger pixel size, smaller array resolution

and most importantly a much higher cost compared with silicon based detectors such as CCD or CMOS (summarized in table 1). etc., will drive SNR to a lower number. This study focuses on the methods to improve system level SNR to enable silicon based detector for this application. Figure 5 - Example of visible spectral response on a silicon wafer Figure 3 - Silicon optical properties: spectral reflection, transmission and absorption Figure 4 - (left) Silicon/InGaAs based detector spectral response, (center) traditional silicon based detector, (right) InGaAs detector Table 1 - Summary of current...

Table 2 - Raw images noise summary patterned but defect-free area to generate golden image, and 3) background images from the non-patterned area. Each set of data includes 20 images for system level noise evaluation. At first observation, the images are very dark. Patterns are deeply buried under the grayish background, corresponding to the analysis of figure 5. Figure 6 - Raw BMP images: (left) out of non-patterned sandwiched-silicon wafer; (center) patterned image without defects; (right) patterned image with defects BASELINE NOISE STUDY System level noise is first analyzed based on the above images....

data value), then map data range to 0~255 based on linear modulation of CDF function, a process based on data value show up possibility. Figure 8 compares the raw image to the equalized and normalized image. Very obviously, the image detail visibility is enhanced dramatically, now missing patterns and the irregular defects are clearly seen on the processed image. Additional flat fielding process removes the non-uniform background so the whole image background looks more uniform for human vision as shown in figure 9. A square matrix with all pixels equal to 1 is generated as a low pass filter....

Based on the analysis in figure 5, in theory the best SNR that can be achieved is 7.6%: useful signal strength is 2.4%, about 13 times less than the top surface reflection. Even though the statistical number shows low system noise, the defect absolute contrast is still too small compared to system noise, which results in the defect being visible but not detectable, as shown in figure 10.1. Only when the absolute defect contrast is larger than the system noise floor, is the defect visible and detectable. When the microscope is focused at the double wafers’ sandwiched inner surface, the top surface...