1.

Introduction

For many years, ongoing requirements for increased computational power at increasingly higher device-packaging densities were accomplished by shrinking the sizes of the basic devices themselves to ensure higher degrees of integration. In the last few years, three-dimensional (3-D) integration has emerged as a complementary method to feature-size scaling to achieve the performance improvement, in which integrated circuits are stacked together in order to improve power consumption, reduce resistance–capacitance delays, decrease device form factor, provide heterogeneous integration, and increase bandwidth, allowing the most efficient process technologies to be used for the various types of devices. One embodiment of this architecture stacks multicore central processing unit (CPU) units with memory devices in the same package.

Several process flows have been proposed for manufacturing 3-D integrated circuits, commonly referred to as via-first, via-mid, and via-last.^1,2 Each of these 3-D technologies requires new wafer-level process technologies to be added to the process flow. These process steps include fabrication of through-silicon vias (TSV), wafer thinning, and either temporary or permanent wafer bonding.

2.

Instrument Description

This article reports on a technique under development using infrared (IR) microscopy^4,5 to identify and characterize the voids. The absorption edge of silicon is approximately 1 μm, which means that the IR wavelengths are necessary for the wafers to be transparent and hence for the voids to be seen by the microscope.⁶ Since this technique uses reflected IR microscopy, it is expected to be especially useful for temporary bonding processes where the interface can be imaged through the unpatterned carrier wafer.

The technique is particularly interesting given its relatively simple setup and fast detection with reasonable spatial and depth resolutions. As will be shown in this publication, good progress has been made in applying IR microscopy to bonding interface void detection. Good correlation to known void dimensions is demonstrated, and the technique’s limit of detection is explored by varying void diameter, density, and depth. Further work is ongoing to integrate the technique into a fully automated metrology tool.

The work presented in this publication has shown IR microscopy to be a promising technique for detecting and measuring geometries of wafer-to-wafer bonding interface voids. A basic diagram of the instrument can be seen in Fig. 1. The instrument has a broadband IR light source that is incident onto the wafer, through an optical filter and a beam splitter. Optical filtering is performed using a 1-μm high-pass filter in sequence with a 1.3-μm low-pass filter, resulting in a wavelength range of 1.0 to 1.3 μm. For detection, an indium gallium arsenide line camera is used, which has good sensitivity in the 1.0 to 1.7-μm wavelength range. Since silicon is transparent in the IR, the instrument is capable of detecting and measuring voids through the top silicon wafer, which is typically $\sim 775\text{-}\mu \text{m}$ thick. The objective lens assembly contains a lens exchanger, which is fitted with lenses of varying magnifications for resolving different size features. The objectives used for image acquisitions to date have magnifications in the range of $5\times$ to apertures of 0.1 to 0.65, and field of view from 8 to 0.8 mm. The camera pixel size calibration is performed by measuring an object of known size in micrometer to determine its size in pixels. After the calibration has been performed, the lateral dimensions measured by the microscope are directly reported during the measurements. The system includes a wafer flipper, as seen in Fig. 2, in order to enable the inspection of the wafer through the bottom substrate. Handling of the wafer is done with edge gripping, and the wafer stage used is a high-precision $XYZ$ stage.

Fig. 1

Basic configuration of the infrared (IR) microscope.

Fig. 2

Wafer flipper used to facilitate the imaging through the backside of wafers.

The microscope images are acquired using the line camera, which has its detector array oriented perpendicularly to the primary scan direction of the stage. The width of the field being scanned therefore depends both on the size of the line array and the magnification of the lens objective being used. Along the scan direction, the image is created by stitching consecutive line image captures as the wafer is scanned at constant speed under the lens objective. The stage motion controller is used to generate a trigger signal for image acquisition. When the stage has moved a predetermined scan length a trigger signal is sent to the line camera, which is synchronizing image acquisitions with the movement of the stage. This predefined scan length, corresponding with the time between triggers, is set so that the pixel size in the scan direction equals the pixel size in the direction along the line camera. The acquisition of a single line starts when the trigger signal is received by the camera. The exposure time is therefore set to be somewhat shorter than the time required for the stage to travel the distance between two triggers.

The two measurement modes available are full and partial wafer scans. For partial wafer scans, the images are captured from predefined areas of the wafer, which regions may be configured to represent dies on the wafer or to cover areas of the wafer that are prone to have voids in the bonding process. The length of the scan is therefore dependent, on which one of the two measurement modes is being used. In case of full wafer scans, the scan length is determined by the wafer boundaries and is therefore dependent on how close to the center of the wafer it is being acquired, whereas for the partial scan mode it is determined by the width of the fields to be scanned. Since the size of scanned areas is generally larger than the field of view of the line camera, a final image is composed of multiple image scans that are stitched together with a small overlap, ensuring full coverage of the scanned area.

The actual time needed for a full wafer scan is dependent on which magnification the objective lens is being used, since higher magnification results in smaller pixel size and smaller field of view. Using a $5\times$ microscope objective, a full wafer scan takes approximately 10 min, while a full wafer scan using a $20\times$ microscope objective requires a little over 1 h. It is therefore useful to be able to control the magnification with the goal to keep the measurement time as short as possible for a given void feature detection wafer scan.

The depth of field (DOF) of the imaging system is dependent on the magnification lens objective used. For the $5\times$ magnification, the DOF is greater than 100 μm, while for the $50\times$ magnification it is a few micrometers. Since the wafer is supported by its edges, there is some sagging of the wafer toward its center. Experience has shown the DOF to be sufficient to resolve a complete field of view, but re-adjustment of the focus height is needed as the wafer is being scanned, and several types of autofocus systems are currently being evaluated for the final metrology tool.

The basic void detection principle is based on the fact that a reflection can occur at the transition in different refractive indices. In this case, because the void has relatively lower refractive index compared to the silicon dioxide bonding material, it gives a different reflectance than the surrounding material. As a result, the amount of light reflected at the bonding interface is different in areas where the voids are present, allowing the voids to be detected by the IR microscopy instrument. An example of this is shown in Fig. 3.

Fig. 3

Principle of IR microscopy: the mismatch in refractive index due to the presence of a void results in greater reflection.

The horizontal geometry of the voids can be determined directly by measuring the voids seen on the indium gallium arsenide (InGaA)s camera. An example of intensity linescan across a void is shown in Fig. 4. As can be seen, the intensity is around its maximum value for a considerable portion of the void diameter, and there are minima near the edges of the void, which have somewhat lower intensity than the background. To determine the lateral size of a single void, we used a 30% intensity threshold edge-detection algorithm. The results obtained from the experiment using this method were found to be in good agreement with nominal designed dimensions for various void sizes, and good numerical stability was achieved when calculating the diameter of different voids of the same size. The distance found is scaled using the pixel size calibration before reporting the void size as the measurement result.

Fig. 4

Linescans across 275-μm wide void features, showing the minimum intensities near the edges. 40-, 400-, 800-, and 1200-nm void depths from top to bottom (45% of original image scale).

In order for voids to be measureable, they need to be resolvable and detectable. Detecting void defects can be done by inspecting an acquired image for intensity changes that are not part of the intentional pattern of the sample being inspected. In order for a detected void to be resolvable, it needs to be separated in the acquired image from neighboring voids. Examples are included in the experimental results for larger areas of multiples voids that are detectable, but with pitch too small for individual voids to be resolvable or measureable. The measurement algorithm used allows reporting of dimensions and locations for voids that are resolvable and at least three pixels in size.

3.

Sample Description

The wafer set consists of four bonded wafer pairs, which are formed using standard thickness (775 μm) 300-mm wafers that are bonded together using a 5-nm oxide layer, using the process flow as described earlier. As shown in Fig. 5, one of the wafers in each pair was patterned with programmed voids. These programmed voids have diameters ranging from 0.5 to 300 μm. The voids on each of the four wafer pairs were etched to different depths ranging from 40 to 1200 nm. In addition, the voids are present in isolated, semi-dense, and dense formats, as shown in the CAD layout included in the left part of Fig. 6.

Fig. 5

Cross-section of bonded wafers showing etched voids.

Fig. 6

CAD layout and captured image of designed defects using a lens objective with $5\times$ magnification, corresponding to an $8\mu \text{m}/\mathrm{px}$ pixel size.

4.

Experimental Results

The right part of Fig. 6 shows the captured image on a die with approximately 400-nm deep voids. The image capture was performed using a lens objective with $5\times$ magnification or $8\mu \text{m}/\mathrm{px}$ pixel size. In this configuration, the image size obtained by one scan is $3600\times 1024\text{pixels}$. To image one die, five overlapping scans were performed, of which the middle $3300\times 737\text{pixel}$ area was used for each scan. Therefore, the resulting matched image (Fig. 6, right) is $3300\times 3685\text{pixels}$. From the image, all the voids in semi-dense ($5:1$ space:width) and dense ($1:1$ space:width) arrangements are detected, but for smaller void sizes, the individual voids cannot be resolved. From Fig. 7, which is showing images from the same image capture at original image scale, we can see more clearly how the smallest resolvable void depends on density and lateral dimensions. As the voids get smaller and thinner, they become more challenging to detect in isolated arrangement and to resolve in regions with increased void density.

Fig. 7

(a) Original scale images of 40-nm deep voids, captured with a $5\times$ objective: 2.5-, 5-, 10-, and 15-μm lateral sizes from top to bottom. For this void depth, a 10-μm detection limit was observed for isolated areas, while all semi-dense and dense voids can be detected but cannot be individually resolved. (b) Original scale images of 400-nm deep voids, captured with a $5\times$ objective: 2.5-, 5-, 10-, and 15-μm lateral sizes from top to bottom. For this void depth, all voids were detected for isolated, semi-dense, and dense areas but all of them were not individually resolved. (c) Original scale images of 800-nm deep voids, captured with a $5\times$ objective: 2.5-, 5-, 10-, and 15-μm lateral sizes from top to bottom. For this void depth, all voids were detected for isolated, semi-dense, and dense areas but all of them were not individually resolved. (d) Original scale images of 1200-nm deep voids, captured with a $5\times$ objective: 2.5-, 5-, 10-, and 15-μm lateral sizes from top to bottom. For this void depth, all voids were detected for isolated, semi-dense, and dense areas but all of them were not individually resolved.

A summary of the smallest detectable, resolvable, and measurable voids for samples with varying void depths is included in Table 1. For the $5\times$ lens objective measurements, the 40-nm deep voids were observed to have a 10-μm detection limit for isolated areas. As the void depth increases to 400 nm or thicker, the smallest detectable void decreases to 2.5 μm for isolated voids, whereas in case of semi-dense and dense areas the detection limit is 0.5 μm. The smallest resolvable 40-nm deep void lateral sizes were found to be 10 μm for semi-dense and 15 μm for dense areas, while a 5-μm resolution limit was found for all other depths. The smallest automatically measurable void size for 40– to 800-nm deep voids was 25 μm for all regions. As it is seen in Fig. 4, 1200-nm deep voids have lower contrast with respect to the background, which leads to unstable measurements of the size of the 25-μm voids, and therefore a 50-μm measurement limit is recorded for this depth. Using the $50\times$ magnification objective lens ($0.8\mu \text{m}/\mathrm{px}$ pixel size), voids as small as 0.5 μm can be detected. In semi-dense areas, voids as small as 0.5 μm can be resolved, while only 2.5 μm ones can be resolved in dense areas. Automatic size measurements were not performed using $0.8\mu \text{m}/\mathrm{px}$ images.

Table 1

Summary of the smallest detectable, resolvable, and measurable voids for samples as a function of void depth and density.

Lateral void dimension measurement fidelity was evaluated by correlating the lateral dimensions measured by the IR microscopy instrument with the nominal (design) dimensions. As shown in Fig. 8, the results are in good agreement with $R$-square $>0.99$ for all the void depths. From the results, we also note that the slope is generally close to unity, indicating that the instrument’s pixel size calibration is in good agreement with the nominal designed dimensions. One exception is the 1200-nm deep voids, for which the slope is 0.94. This is not yet fully understood. Further work is ongoing to measure the sister wafers using other techniques, which may help to give a better understanding. While the intentional voids greater than 1-μm deep that are etched in silicon for the DOE wafers are interesting for exploring instrument sensitivity, the bonding voids that pose practical issues in the semiconductor fab manufacturing line are typically formed in the bonding interface.

Fig. 8

Correlation between lateral dimensions measured by the IR microscopy instrument versus the nominal dimensions based on the CAD layout. These results are based on data of isolated voids. Measurements on top left voids (and their nearest neighbors) in semi-dense and dense arrangements resulted in same values within measurement error.

In order to further explore the sensitivity to void depth, reflectance simulations were performed for void depths on the DOE wafers (40, 400, 800, and 1200 nm).^7,8 For the simulations, a 0-deg angle of incidence was used with a filmstack from bottom to top as follows: infinite silicon substrate, air (varied thickness), 5-nm silicon dioxide, and 775 μm of silicon. As shown in the simulation results in Fig. 9, reflectance is expected to increase from 40 to 800 nm, but for 1200-nm depth, antireflective behavior is observed with lower average intensity. It should be noted that in this study, the focus has been on voids that are formed in the lower part of the oxide bonding layer. Additional future work exploring the effect of bonding interface void depth on instrument sensitivity could be beneficial.

Fig. 9

Simulation results showing spectral sensitivity to void depth.

Results shown in Fig. 10 indicate that the intensity contrast is sensitive to the void depth, and good correlation is observed. The contrast was calculated as the normalized range in intensity measured across the voids. A future study may be performed to determine whether the measured contrast together with precalculated reflectance versus void thickness data can be reliably used to determine the void thickness.

Fig. 10

Correlation between simulated reflectance and measured contrast.

5.

Conclusions

The IR microscopy has been evaluated for detecting and measuring the dimensions of voids that are formed in the oxide-to-oxide interface of permanently bonded programmed void wafers. Results have demonstrated that the technique has the required sensitivity to detect and measure the isolated and dense voids of horizontal dimensions varying in range from submicron up to hundreds of micrometers and depths varying from 40 up to 1200 nm. The capability to acquire the images, inspect them for void defects, and measure the found voids with reasonable speed makes the technique a good solution for high-volume manufacturing implementation in semiconductor fabs.

The IR microscopy instrument is currently being integrated into a fully automated platform in order to build a tool suitable for tier 1 fab production implementation. The method is expected to have immediate application to temporary bonding applications, where the imaging can be done through the carrier wafers. Additional work will be needed to investigate the applicability to permanent bonding applications, where the presence of surface metallization and TSVs (for via-first and via-mid processes) may prevent imaging of the bond plane. Work is also being pursued to explore whether combining IR microscopy with other techniques, such as photoluminescence or model-based Fourier transform infrared (FTIR) reflectometry,⁹ will be beneficial for improved inspection and metrology capability in a manufacturing metrology tool. Future work to take advantage of algorithms previously developed is desirable to more accurately report dimensions.^10,11