Comparison of surface energies calculated with analytical and FEM models
The apparent surface energies of two surface step test structures and two recess test structures are plotted in Fig. 7. The heights and depths of the surface steps and recesses are 10 nm and 15 nm. Generally the apparent surface energies increase with decreasing and because the required deformation energy increases when the deformed area decreases at constant step height or recess depth. That is to say, high surface energies induce stronger deformation.
Calculated apparent surface energy γ as a function of (a) the length of the unbonded region near the surface steps and (b) the width of the unbonded rim in the circular recesses with b = 200 µm. The height of the step h (a) and the depth of the recess h (b) are varied. The thickness of both wafers are 675 µm. The curves were calculated using the present analytical model (Eq. (4)), the analytical model by Martini et al. (Eq. (5)), and the FEM simulation software together with either Eq. (6) (a) or Eq. (1) (b).
In both cases the present analytical models, Eqs. (4) and (9), and the FEM simulations agree well when the dimensions of the unbonded regions, and , are larger than 10 mm. At shorter unbonded regions the present analytical models, Eqs. (4) and (9), yield orders of magnitude higher surface energies than the FEM simulations. The reason for this is that at small dimensions ( and less than 1 mm) the analytical models based on the thin plate theory overestimate the mechanical energies needed in the deformations. At small dimensions, in fact, the thin plate theory is not valid because the thickness of the wafer should not be more than about 1/4 of the least transverse dimension of the deflected structure.25 Most of the practical geometries and the test structures utilized in this work violate this assumption as well. In spite of this, the analytical formulas provide still insightful information how the surface energy depends on the geometrical and material parameters.
The analytical model by Martini et al., Eq. (5), agrees with the FEM simulations when (see Fig. 7a), but yields 1–3 orders of magnitude lower surface energies when . This difference is presumably due to local high stress points at the shoulder corners of the recess test structures. These high stress points, which were seen in the FEM simulations, increase total strain energy U of the structures, thus contributing to higher apparent surface energies. The analytical models are not able to describe these high stress points, but all of them are able to describe well the h-dependence, γ∝h2, in the whole range of and .
In summary, Fig. 7 shows that the FEM and analytical models of this work agree very well when the dimensions of the test structures are large. When and are less than 1 mm, however, only the FEM models can be used since the present analytical models are not valid in this range.
Dependencies of apparent surface energies on bond test structure geometries
The calculated apparent surface energies of the recess and mesa bonding test structures are plotted in Fig. 8a and b as functions of the width of the unbonded rim. The graphs show that these test structures can be used for measuring the apparent surface energies in the range 10− 4–10− 2 J/m2.
Calculated apparent surface energy γ as a function of the width of the unbonded rim (a) in circular recesses with various values of a and (b) around circular mesas with various values of b. The depth of the recessed h and the height of the mesas h are 10 nm. The curves were calculated using the FEM simulation software and Eq. (1) with wafer thickness of 675 µm.
The apparent surface energies of the surface step bonding test structures are plotted in Fig. 9. The dynamic range of the surface steps extends to as high as 10 J/m2. Across the range of interest the h-dependencies of the curves in Fig. 9b are extremely close to the h2-dependency predicted by the analytical models, Eqs. (4) and (5). This was tested by fitting the curve γ = αh2, where α is a constant, to the data. The relative root-mean-square (RMS) errors of the fits were only 1–3 ppm. Similar results were obtained also for the mesa and recess structures.
Calculated apparent surface energy γ as a function of (a) the length of the unbonded region near a surface step and (b) the height h of the surface step. h is varied from 10 nm to 19 nm in (a) and from 60 µm to 140 µm in (b), respectively. The curves were calculated using the FEM simulation software and Eq. (6) with wafer thickness of 675 µm.
Effect of temperature and dilution of SC1 on fusion bonding
The effect of standard and dilute SC1 was studied at 45 °C and 65 °C. In general, the use of the lower dilution improved bonding quality at both temperatures, and higher bonding quality was obtained at the lower temperature regardless of the dilution. Below we focus on the details of the extreme cases.
The unprocessed SAM images of fusion bonded wafers with standard and dilute SC1 cleaning treatments at 65 °C and 45 °C, respectively, are shown in Fig. 10. Most of the large square recesses in the dilute SC1 wafers are bonded in Fig. 10b, whereas in the standard SC1 wafers most of them are not bonded in Fig. 10a. In addition, the narrower unbonded rims in the inset of Fig. 10b are clearly visible.
SAM images of fusion bonded wafers I.S and I.D with 11 nm deep recesses shown in Fig. 4a. The wafers were cleaned in (a) standard SC1 at 65 °C and (b) dilute SC1 at 45 °C. The insets show magnifications of the wafer centers.
Maps of the widths of the unbonded rims and the corresponding apparent surface energies obtained from the SAM images using the procedure explained in the experimental section are shown in Fig. 11. Fig. 11 shows that the wafer pair activated in dilute SC1 has much narrower unbonded rims and the corresponding higher apparent surface energies than the wafer pair activated in standard SC1. Also, the variation of the apparent surface energy across the wafer is more uniform in the case of dilute SC1.
Measured widths of the unbonded rims in the 2000 µm wide and 11 nm deep recesses square recesses in the wafers cleaned in (a) standard SC1 at 65 °C and (b) dilute SC1 at 45 °C. The data was obtained by deconvolution and analysis of the SAM images of Fig. 10. The corresponding apparent surface energies are also shown. Completely unbonded recesses are shown in black. (a) 15.5% and (b) 86.4% of these recesses were bonded. The average unbonded rim widths were (a) 134.7 µm and (b) 98.31 µm.
In experiments II.S and II.D (see Table I) the wafers were investigated with SAM before and after the annealing steps. These SAM images were closely similar to the SAM images obtained after the annealing. This supports the fact that these kinds bond test structures characterize the bonding interface at the time of wafer contact.12
SAM images of the recess and mesa test structures of Fig. 4b and c on fusion bonded wafers are shown in Fig. 12, where differences in the dimensions of the unbonded regions are clearly seen. The average widths of the unbonded regions and unbonded rims are plotted in Fig. 13 as functions of the surface step height, mesa height, and recess depth, respectively. The calculated surface energy curves in Fig. 13 were obtained from the FEM simulations by assuming and .
SAM images of bond test structures of (a,b) Fig. 4b and (c,d) Fig. 4c on fusion bonded wafers with 12 nm deep recesses (experiments II.S and II.D). The wafers were cleaned in (a,c) standard SC1 at 65 °C (II.S) and (b,d) dilute SC1 at 45 °C (II.D).
Measured wafer-level values of (a) average of a surface step (from 2000 µm wide square recesses), (b) average of a 600 µm circular mesa, and (c) average of a 2000 µm circular recess as a function of h. Two different SC1 treatments are compared (experiments II.S and II.D). Lines are calculated constant apparent surface energies. Each data point represents average of (a) 41, (b) 72, and (c) 41 test structures on a single wafer.
Fig. 13 shows that the surface step, mesa, and recess models produce slightly different apparent surface energies from the experimental data. The surface step model estimates the highest apparent surface energies: 12–30 mJ/m2 for dilute SC1 (45°C) wafers with step height of 10–19 nm, and 13 mJ/m2 for standard SC1 (65°C) wafers with step height of 12–16 nm. The mesa and recess models estimate lower values: 6–14 mJ/m2 for dilute SC1 (45°C) wafers with h of 10–19 nm, and 7–9 mJ/m2 for standard SC1 (65°C) wafers with h of 12–16 nm.
Since the apparent surface energies of the wafers with the same composition and surface treatment should be the same, these results suggest that the present models are not fully applicable in the present experiments. The most probable reason for this is the fact that the present models do not take the 500 nm thick silicon oxide layer into account. However, the present models are still useful in giving estimations of the surface energies. Addition of the oxide into the models needs careful modeling of the stresses originating from the oxidation process. Therefore, further information on these stresses and the details of the oxidation process are needed.
As it has been already shown some of the recesses are bonded and some are not. The bonding probability of a recess depends on the depth, the lateral size, and the shape of the recess as well as the activation treatment of the surfaces. The measured bonding probabilities of circular and square recesses (see Eq. (10) in the experimental section) as functions of the area and the depths of the recesses are shown in Figs. 14 and 15. The bonding probability decreases with increasing recess depth and increases with increasing recess size. The shape has a minor effect on the probability, at least in the case of squares and circles. The surface treatment has a huge impact on the bonding: The bonding probability of a 2 mm wide and 12 nm deep square recess is 71.8 % on a wafer cleaned in dilute SC1 at 45 °C and 12.7 % on a wafer cleaned in standard SC1 at 65 °C (see Fig. 14).
Measured wafer-level bonding probabilities of circular and square recesses as functions of the area of the recess and the varying recess depths and SC1 treatments (experiments II.S and II.D). The bottom graph shows a magnification of the upper graph. Lines guide the eye.
Measured wafer-level bonding probabilities of circular and square recesses as functions of the depth of the recess h and varying diameters of the recesses and SC1 treatments (experiments II.S and II.D). Lines guide the eye.
This statistical problem of recess bonding is similar to the yield problem in integrated circuit (IC) manufacturing: In IC manufacturing local, stochastic defects on wafers cause failures of some of the manufactured ICs. In this work, local, stochastic features on wafer surfaces cause bonding of some recesses. Therefore, yield models28,29,30 can be used to describe the probability of a bond initiation site being found in a recess. It should be also noted that kinetic models cannot be used here, because there is no equilibration process: The bonding of the recesses is a fast process which stops after the bonding front has passed the recesses. The early phases of the wafer bonding process can be described by models initially developed for stiction in MEMS.31,32 In general, the interaction between the bonding surfaces is determined by capillary forces, van der Waals forces, electrostatic forces and forces related to hydrogen and solid bridging.31,32 It is very unlikely, however, that capillary forces play a role in vacuum bonding of recesses. In addition to the environmental conditions such as temperature, the roughness of the bonding surfaces has profound effect on the bonding process.
In the yield models, the yield Y of the fabricated devices is limited by defects. In the present problem, however, there are local disturbances on wafers, which cause the bonding of recesses to occur. In the context of a yield model, the defects themselves cause the bonding. Therefore, the probability of a recess to bond can be written as 1 − Y.
We analyzed the data of Fig. 14 with two general models, the negative binomial model28,29 and the variable defect size (VDS) model.30 These yield models are able to explain the behavior of the 10 nm deep recesses with areas higher than 1.5 mm2. The VDS model suggest that the bonding is caused by a point-like sources with a density of 0.9 mm− 2. These models and the other common yield models,28,29,30 however, predict much higher probabilities than observed when the recess area is smaller than 1.5 mm2. This discrepancy is even higher with deeper recesses and the standard SC1 cleaning. The discrepancy is most likely because these models do not take the energy needed to bond into account. In general, the density of the initial energy needed for the recess to bond increases with decreasing recess area simply due to the fact that the bonding of smaller recesses must begin with smaller value of dcirc (see Fig. 7b). Presumably, the limited availability of energy during the bonding process reduces the bonding probability if the required energy is high.
Fig. 16 shows the dependence of the width of the unbonded rim of the 2000 μm square recesses on the bonding probability of the same recess. There is a clear dependence between these two: When the unbonded width is small, the bonding probability is large and vice versa. Since the width of the unbonded rim is inversely proportional to the apparent surface energy, this suggests, rather obviously, that the bonding probability of a recess increases with increasing apparent surface energy of the bonding surface. This effect is also clearly visible in Fig. 10.
Measured width of the unbonded rims of the 2000 μm wide and 12 nm deep square recesses on bonded wafers treated in SC1 with various dilutions and temperatures (experiments III.S1, III.S2, III.D1, and III.D2) as functions of the measured wafer-level bonding probabilities. The corresponding apparent surface energies are also shown. Each data point corresponds to a single test structure on an individual wafer. The unbonded rims of typical recesses were measured manually with a high-resolution SAM scan.
The experimental data in Fig. 16 shows that the apparent surface energy of the SC1 treated surfaces decreases in the order 1) dilute SC1 at 45oC, 2) standard SC1 at 45oC, 3) dilute SC1 at 65oC, and 4) standard SC1 at 65oC. The data suggests that as the strength of the SC1 mixture increases, the apparent surface energy decreases. These results can be explained by the fact that the surface roughness increases as the NH4OH concentration in the SC1 solution increases.11
The SC1 results are summarized in Fig. 17, which also shows the results obtained with the COT. Although COT measures the surface energy as the mechanical energy needed to pull the bonded wafers apart, both apparent surface energy (Fig. 17b) and the recess bonding probability (Fig. 17c) correlate well with the final bond strength of the SC1-treated wafers measured by COT (Fig. 17a). Fig. 17 shows also the results from the plasma-activated wafers, which are discussed in detail in the next section. The apparent surface energy and the recess bonding probability are both measures of the bonding surfaces at the time when the wafers are first brought into contact. The further bonding mechanisms taking place during annealing determine the final bond strength. Because the mechanisms of plasma bonding differentiate remarkably from those in the fusion bonding, the fusion and plasma bonding results in Fig. 17 cannot be directly compared.
Comparison of results from fusion bonded and plasma-activated wafers: (a) surface energies measured using the crack-opening test (COT), (b) the apparent surface energies obtained from the bonding test structures (Figs. 11 and 19), and (c) the bonding probabilities of 11 nm (SC1) and 12 nm (O2 and N2 plasma) deep recesses (Figs. 11 and 19). The values are measured averages of (a) 3 samples from a single wafer pair and (b,c) 121 test structures on a single wafer. The standard SC1 cleaning was performed at 65°C and the dilute SC1 at 45°C. The COT surface energies of the SC1-treated wafers were measured after annealing at 1100°C for 2h, whereas the O2 and N2 plasma-activated wafers were measured directly after bonding at room temperature (RT) and after annealing at 400°C for 1 h.
Overall, the estimation of surface energy based on the measured bonding probability is, indeed, a fast and simple way to compare cleaning treatments, but does not allow mapping of surface energy and cannot be quantified without reference samples. The results imply that the surface energy at the time the wafers are brought into contact determines also the final bond strength of SC1-treated wafers which is obtained after annealing.
Comparison of nitrogen and oxygen plasma activated samples
Examples of SAM images of plasma activated wafers are shown in Fig. 18. Maps of the widths of the unbonded rims and the corresponding apparent surface energies obtained from SAM images (see the experimental section) are shown in Fig. 19. Compared to the recesses on the fusion bonded wafers (see Fig. 11) the number of bonded recess on the plasma activated wafers increases toward the edges of the wafers. This could be caused by the spatial variation of the plasma in the processing chamber.
SAM images of bond test structures of Fig. 4a on plasma bonded wafers with 12 nm deep recesses (experiments IV.N and IV.O). The wafers were activated in (a) nitrogen (IV.N) and (b) oxygen plasma (IV.O).
Measured widths of the unbonded rims in the 2000 µm wide and 12 nm deep square recesses in the wafers activated in (a) nitrogen and (b) oxygen plasma (experiments IV.N and IV.O). The corresponding apparent surface energies are also shown. Completely unbonded recesses are shown in black. (a) 28.5% and (b) 51.7% of these recesses were bonded. The average unbonded rim widths were (a) 100.2 µm and (b) 95.3 µm.
Fig. 17 shows the comparison of the SC1 and plasma activated wafers as well as the bond strength of the plasma-activated wafer measured with COT both before and after annealing. Both plasma-activated wafers have apparent surface energies near 20 mJ/m2, which are in agreement with the values 40–75 mJ/m2 reported by Bodner et al.33 The results in Fig. 17 suggest that the surface energy and the resulting interaction of the bonding surfaces is stronger in O2 plasma activated wafers than in N2 plasma activated wafers and even SC1-treated wafers, but this does not turn into high bond strength during annealing. In the case of nitrogen plasma, strong final bond strength is obtained during annealing in spite of weaker interaction. Compared to the fusion bonding of the SC1-treated wafers, differences between the interaction of the bonding wafers can also explain the fact that the apparent surface energy and the recess bonding probability of the plasma-activated wafers are only slightly correlated with the COT surface energies measured at RT and not correlated with the COT surface energies measured after annealing.
Whereas the SC1-treated samples allow a straightforward interpretation of apparent surface energy in terms of surface roughness, the plasma activated wafers are not so simple. As Fig. 17 shows the apparent surface energy of the N2 activated surface is comparable to that of the SC1 activated surfaces. The O2 activated surface is even higher. This contrasts with the typical expectation, confirmed here as well, that the bond strength of the N2 and O2 activated wafers after annealing are often lower than that of the SC1-activated fusion bonded pairs. Furthermore, O2 activated pairs should be weaker than N2 bonded pairs. The discrepancy in the behavior of SC1 activated and plasma activated surfaces may lie in the nature of the surfaces created.
The surface energy of a solid has both polar and non-polar components. A perfect, defect free oxide surface comprised of bridging oxygen atoms is effectively hydrophobic. A fresh, thermal oxide provides an example of an oxide surface with a relatively high contact angle (hydrophobic)34 with a surface covered mainly in siloxanes.35 Introduction of defects will increase the surface energy toward hydrophilic behavior. Non-bridging oxygen atoms have been modeled on such surfaces.36 Wendt and co-workers37
Wafer bonding is as a key technology for microelectronics and micro-electro-mechanical systems (MEMS) for creating three-dimensional (3D) structures and hermetic packaging. Several bonding techniques have been developed in the last several decades, including silicon fusion bonding, anodic bonding, solder bonding, adhesive bonding, and eutectic bonding [1,2,3]. Compared with other bonding technologies, Au/Si eutectic bonding benefits from a low eutectic temperature (363 °C), being non-sensitive to the roughness of bonding surface and particles, having good mechanical stability, and compatibility with aluminum interconnect . Therefore, it is considered the most important approach to establishing strong bonds and hermetic seals, and is widely used in the fabrication of MEMS devices, 3D interconnects, and wafer-level vacuum packaging [5,6]. As such, Au/Si eutectic bonding has been investigated extensively. Maryam Abouie et al.  found craters on the bonding interface and attributed them to the anisotropic and non-uniform reaction between Au and Si. Michael David Henry and Catalina R. Ahlers  studied the effect of platinum diffusion barriers and the metallization process in Au/Si eutectic bonding. They concluded that the barrier broke down at approximately 375 °C and resulted in uncontrolled stoichiometry variations and the creation of micro voids. Therefore, the bond strength and hermeticity were reduced, but the alternative material was not presented. Errong Jing et al.  investigated the interface of Au/Si (100) eutectic bonding with an infrared (IR) microscope and related their observations to the bond strength. The results showed that a strong bond had many square black spots in the IR images, whereas a poor bond had fewer or no square black spots. This method provided a convenient method using an IR microscope to evaluate bonding quality without destroying the bonded pairs. Furthermore, the method for improving the reliability of Au/Si eutectic bonding with either an Mo-buffered layer  or localized bonding  was also studied. However, there have been few reports about the influence of contact force and the heating/cooling rate on Au/Si eutectic bonding quality. In addition, new stacks of metal layers are still needed for simplifying the fabrication process and enhancing the bonding quality.
In this work, we developed a simple Au/Si eutectic bonding process that enabled high bonding strength and a uniform bonding interface for hermetic packaging. The influences of process parameters (such as bonding temperature, contact force, and heating/cooling rate) on the bonding strength and bonding interface were studied. Pyrex 7740 glasses with patterned Cr/Au, Ti/Pt/Au, or TiW/Au layers were bonded to Si wafers. The bonding interfaces were observed and analyzed with an optical microscope, a scanning electron microscope (SEM), and a related energy dispersive spectrometer (EDS). The bonding strength and bonding yield were evaluated with a tensile pulling test and a dicing experiment, respectively. Finally, a wafer-level vacuum packaging structure was designed for an MEMS accelerometer, which achieved hermetic packaging and a metal wire interconnection with a single bonding process, and the performances of the accelerometer were examined.
2. Au/Si Eutectic Bonding
Au/Si eutectic bonding is a kind of intermediate layer wafer bonding that utilizes the low melting temperature of an Au/Si eutectic alloy. The principle of Au/Si eutectic bonding is quite similar to soldering. A liquid phase is formed when the temperature exceeds the eutectic point, and Au and Si are then mixed together with a certain ratio. Subsequently, the material mixtures are solidified regardless of whether the temperature decreases below the eutectic point or whether the ratio of the concentration leaves the liquid area. The reaction starts with atomic contact of the partners; with an increase in temperature, the diffusion of Au and Si occurs. As soon as the eutectic mixture reaches the eutectic composition, the liquid phase begins, and further mixing and diffusion processes are accelerated. From the phase diagram of Au/Si , it can be seen that the melting point of the Au/Si eutectic alloy is around 363 °C, which is much lower than that of the melting points of Au (1063 °C) and Si (1412 °C) respectively. The eutectic composition is 19 atom % (3 wt %) Si and 81 atom % (97 wt %) Au.
3. Experimental Details
3.1. Wafer Preparation
Four inch double-polished n-type (100) Si wafers with resistivity around 2–4 Ω·cm (Dingjing Co., Ltd., Luoyang, China) and Pyrex 7740 glass wafers (Taikunisi Co., Ltd., Suzhou, China) were used for bonding experiments. Figure 1 illustrates the Au/Si bonding structures studied in this work. The Pyrex 7740 glass was immersed in H2SO4:H2O2 (3:1) at 120 °C for 10 min to remove organic particles. The bonding layer of Cr/Au, Ti/Pt/Au, or TiW/Au were deposited on the glass substrate in sputtering system FHR-MS100 × 6 (FHR Anlagenbau Gmbh, Dresden, Germany) with a base pressure of 1 × 10−6 mbar. The bonding structures on the Si wafer were fabricated by a standard wet etching process.
3.2. Bonding Process
Before bonding, the Si wafer with a bonding structure and the Pyrex 7740 glass wafer with a deposited metal layer were wet-cleaned with fuming nitric acid (a kind of concentrated nitric acid with the nitric acid content of 90–98%) at 100 °C for 5 min to remove organic contaminations. The wafers were then rinsed in DI water and dried with nitrogen. In addition, the Si wafer was dipped in 1:100 HF for 10 min in order to remove the native oxide layer on the Si surface and ensure small surface roughness simultaneously. Immediately after DI water rinsing and nitrogen drying, the wafer pairs were aligned by a mask aligner MA6/BA6 (Karl Suss, Munich, Germany) and then loaded into a Karl Suss vacuum bonder SB6e (Karl Suss, Munich, Germany). After the N2 purge, the bonder was pumped to a base vacuum of 5 × 10−5 mbar, and then ramped to the bonding temperature. The eutectic bonding was carried out at 380–400 °C with an applied force of 2000–5000 N, and then slowly cooled down to room temperature.
3.3. Bonding Quality Evaluation
The bonded wafer was cut into 7.14 mm× 7.14 mm pieces at a speed of 1 mm/s with the dicing saw DS616. The valid bonding areas are 5.49 × 10−6 m2 of a single chip and 8 × 10−4 m2 on the entire wafer. The dicing results can be used to initially evaluate the bonding quality: the separated pieces are considered to be a failed bonding, and the pieces staying together represent successful bonding. Bonding yield is defined as the percentage of successfully bonded pieces on each wafer. After the dicing test, surviving bonded pieces went with a tensile pulling test to determine bonding strength. A force was continuously and perpendicularly applied to the bonding interface until the sample was broken. The bonding strength was determined by dividing the maximum force to the bond area. A similar behavior was detected by razor blade tests for preliminarily estimating bonding strength. A thin razor blade was inserted between the bonded pairs from the wafer edges, and the separated surfaces were inspected with an optical microscope and analyzed with an energy dispersive spectrometer (EDS, ZEISS, Jena, Germany). Either the silicon broke due to a high bonding strength or delamination at the Au/Si interface occurred due to a weak bonding strength. The fracture surfaces were also examined by a scanning electron microscope EV018 (SEM, ZEISS, Jena, Germany).
4. Results and Discussion
Au/Si eutectic bonding is an inter-diffusion process between Au and Si, and heat plays a major role in this process. When bonding temperature exceeds the eutectic point (363 °C), Au and Si are in contact with each other, and the liquid phase alloy with eutectic composition can be formed during the diffusion process. As time goes on, the liquid phase layer becomes thick. During the subsequent cooling process, the liquid phase continuously alternates with these two kinds of metals, each of which is grown, crystallized, and precipitated on the basis of its original solid phase, respectively. Thus, the eutectic alloy between the two materials binds the two wafers together tightly.
The influence of bonding temperature on Au/Si eutectic bonding is examined and analyzed. With the bonding temperature of 380 °C, the bonding strength is 6.7 MPa. It is so weak that the bonded wafers are easily separated during the blade test. Observing the separated surfaces, a slight in-depth diffusion of silicon into gold already occurred, and some silicon particles are seen on the separated surface, as shown in Figure 2a. However, the bonding quality is insufficient for the further applications, especially for vacuum packaging. Thus, the bonding temperature has been raised to 400 °C, and the measured bonding strength is 66.8 MPa. Big pieces of silicon are peeled off from bulk silicon after the blade test, which indicates that the bonding strength is higher than the bulk Si, as shown in Figure 2b. It can be concluded that a high bonding temperature of 400 °C is needed for high bonding strength. According to S. Lani et al. , the high bonding temperature is mainly needed due to the impurity elements in the bonding materials (such as oxygen and hydrogen), or the temperature non-uniformity on the whole wafer, which causes a partial reaction between Au and Si. In addition, the difference between the setting temperature of SB6e and the actual temperature is another factor for a higher bonding temperature. As expected, the bonding strength increases with a further increase in bonding temperature, which is shown in Table 1 (Samples 3 and 8). From Equation (2), a higher bonding temperature results in thicker Au/Si alloys, so the bonding strength increases. However, a high bonding temperature always causes metal wire melting or performance declining for temperature-sensitive devices . Therefore, the optimized bonding temperature is 400 °C.
4.2. Heating/Cooling Rate
The influence of temperature heating/cooling rate on successful eutectic bonding was also investigated. Figure 3a exhibits a failed bonding at high heating/cooling rate of 15 °C/min. The bonded wafers were easily separated by the blade test from the bonded interface. Similar to the observation by Bokhonov et al. , some dendrite crystallizations were seen on the separated Au surface. The component elements on the surface were analyzed via EDS. Results show that the chemical elements were as follows: Si (29.82 wt %), Au (52.72 wt %), Ti (2.8 wt %), and W (14.67 wt %) at Point 1, as shown in Figure 3b, whereas the chemical elements of Point 2 are as follows: Si (10.87 wt %), Au (69.46 wt %), Ti (3.43 wt %), and W (16.24 wt %). Both analyses ignored the influence of carbon (C) and oxygen (O) elements. A small amount of silicon at Point 2 was introduced by the glass substrate. The increase in Si content at Point 1 indicates that a liquid Au/Si alloy was formed. However, the bonding strength is weak because of the formation of dendrites. Au/Si eutectic bonding is a process of liquid phase transformation into a solid phase, and the crystallization occurred during the solidification process. With a high heating/cooling rate, the atoms in the liquid phase can diffuse easily, whereas the atoms in the solid phase are strained to spread due to a lower diffusion rate. Thus, the composition of the later crystal is different from that of the previous crystal, resulting in the formation of dendrites. The dendrites will cause the crystallographic direction to deviate from its previous growing direction and will hinder further solidification of the Au/Si alloy. As a result, a stable and continuous Au/Si alloy cannot be formed. On the other hand, a low heating/cooling rate prolongs the solidification process, indicating a longer liquid alloy time. Therefore, the liquid Au/Si alloy is easily spreads to a non-bonded region (called metal squeezing), especially at high contact force, which will be discussed in the next section. Furthermore, the electrical connection of devices can be damaged because of a serious metal squeezing.
The optimized temperature curve of Au/Si eutectic bonding is shown in Figure 4. First, the samples are heated up from room temperature to 300 °C with the maximum power of SB6e. Then, the temperature rises to 350 °C and remains constant for 10 min for uniform heating of both substrates. Subsequently, the samples are heated to 400 °C with a heating rate of 5 °C/min, which is maintained for 20 min. Finally, the cooling rate is controlled to 5 °C/min for the full solidification of Au and Si. The contact force is applied as the temperature above 300 °C, and the samples are heated from both the top and the bottom at the same time.
4.3. Contact Force
Meanwhile, the influence of the contact force on bonding quality is also examined, as shown in Table 1 (Samples 2, 3, and 6). It can be seen that the bonding strength increases from 37.5 to 66.8 MPa, as the contact force varies from 2000 to 3000 N. On the one hand, high contact force can overcome the non-planar and rough surface, ensuring good contact of Au and Si surfaces. On the other hand, a high contact force is beneficial for the disruption of a native oxide layer at some spots . Thereby, more exposing local sites in the underlying Si diffuse into the Au. The Si on the direct contact region will dissolve into the Au and diffuse rapidly along the grain boundaries of the Au film, so the bonding strength will increase. However, with the further increase in contact force to 5000 N, the bonding strength decreases to 41.4 MPa. This is because a high contact force results in a metal squeezing out of the interface, causing a poor interface layer uniformity, and craters appear at the interface, as shown in Figure 5. Nevertheless, the bonding interface is uniform and no metal is squeezed with a 3000 N contact force.
Furthermore, observing the deviation of the tensile strength in Table 1