A self-assembled covalent nanoglue

Most semiconductor processing bonds used now [1] are either Cu-Cu thermal compression bonds formed at 300 °C–450 °C, sometimes clamped by electrostatic forces [2], or Si-fusion bonds, formed at up to 400 °C. These high temperatures may impact the moieties on the samples and will impact alignment due to thermal expansion [3]. Temperatures at or below 200 °C are preferred [3] to maintain alignment of the materials, although some types of alignment can tolerate 300 °C [4]. Efforts to produce lower temperature bonds, such as using rapid annealing at 290 °C, can call for extreme measures such as high vacuum and require particular surface chemistry [5]. Nano-glues are characterized by very thin extent in the direction perpendicular to the bond, and permit intimate contact of the two surfaces to be bonded. This implies minimal disturbance to propagation of heat or traveling waves, since the layer is thin enough to be bridged by evanescent coupling, similar to frustrated total internal reflection or low temperature 'thermal waves [6].' Prior nanoglues have been proposed [7], but it was demonstrated only to glue an evaporated layer onto the surface of a substrate and not to bond together two wafers. Another nanoglue for deposited layers was used in a sodium battery [8]. Other efforts have used hydroxyl-terminated small TiO₂ particles [9] or sol-gel SiO₂ [10, 11] to glue TiO₂ nanoparticles onto a surface. The use of SAM layers into a standard Cu-Cu wafer bonding technique has been found to improve the electrical properties [12–15]. A review of bonding techniques has been given [16], including reliability issues that we address.

Our aim is to create a bond that is very strong, reliable, and can be incorporated into standard semiconductor processes. For strength, we use covalent bonds including attachment to the surfaces, which is enabled by silane-group-based self-assembled monolayers. Reliability is aided by being very thin, so thermal matching is not required. For long-term reliability and robustness, especially when combined with an alignment process, a nanoglue should (1) not require high temperature to form. Our best process uses 200 °C, (2) be able to withstand high temperatures later if required, as covalent bonds will, (3) no produce byproducts, especially acids or large molecules that might cause corrosion or bubbles to form, (4) not form until initiated, so that fine positioning is possible, as ours is with heat (but flat, clean silicon on silicon is not), (5) applicable to any surface, and (6) be available as a temporary, removable bond or as a strong, covalent, permanent bond.

Most current semiconductor processing bonds used now [1] are either Cu-Cu thermal compression bond (300 °C–400 °C), sometimes held by electrostatic forces [2], or Si-fusion bonds (up to 400 °C). These high temperatures may impact the moieties present on the samples and will impact alignment due to thermal expansion [3]. There is a possibility that some of this can be mitigated with negative thermal expansion layers [17], but the fabrication would be complex. Temperatures at or below 200 °C are preferred [3]; although some types of alignment can tolerate 300 °C [4]. Efforts to produce low temperature bonds (rapid anneal at 290 °C) can call for extreme measures such as high vacuum and limited surface chemistry applicability [5]. Other special purpose bonds can be produced with coordinated polymer stacks, although they need to have a solvent dried out [18], or superhydrophilic nanoglues to bond metal hydroxides to carbon-materials [19]. Our nanoglue is also directly compatible with a fluid whole-wafer alignment method [20]. We demonstrate wafer bonding with a nanoglue layer that uses a thin, permanent layer between the surfaces, applicable to wafers, in contrast to the other efforts referenced above. The nanoglue has covalent bonds from one surface to the other, providing a bond stronger than the substrate materials and thereby providing a truly intimate coupling. Self-assembled monolayers (SAM) on the two surfaces are modified to provide complementary terminations to allow covalent bonding. When pulled apart, covalently bonded wafers either break the silicon underneath or cannot be pulled apart. Some of our intermediate layer samples did not covalently bond for the entire wafer separation. These showed a temperature-dependent debonding force, and could be used for temporary bonding applications.

The idea of a covalent nanoglue bond is realized through the covalent bonding of molecular monolayers that have previously been self-assembled with covalent bonds to two wafers. The most direct way to achieve this covalent glue is the formulation of a peptide or nylon-like bond between the functional groups of the self-assembled monolayers deposited on the oxide layers of prepared silicon substrates. This 'bare' bond would be ideal, because it would be the shortest and probability the strongest bond. A direct, permanent bond would be strong enough to withstand both mechanical stress and heat of, for example, a 3D stack of bonded wafers, with strength similar to or exceeding that of the wafers themselves. The ability to bond SAMs to SAMs would also streamline production of the 3D package, eliminating the need for an added adhesive; since, once alignment is achieved, bonding the wafers together could take place without additional lithography or deposition steps. Finally, using this type of bonding would improve thermal conductivity within the unit, because the adhesion takes place between a very thin layer of integral components of the unit, rather than through additional layers of intermediary materials, which often do not have good thermal conductivity.

Silicon wafers are cleaned in a piranha solution of 2:1 H₂SO₄ (~34 N) and 30% H₂O₂ for 1 h. The H₂SO₄ is added last, slowly. After cleaning, samples are rinsed twice with deionized water. Finally, the wafer is spun dry on a spin coater at 3000 rpm while spraying with HPLC grade isopropyl alcohol. It is then spun for an additional 30 s to completely dry the surface. The SAM layers are grown in a dry box using hexadecane as a dry solvent. The SAMs used to create the nylon bond are 10-undecenyltrichlorosilane and (Aminoethyl aminomethyl) Phenethyl trimethoxysilane. The latter terminates in an amine group. The former terminates in a vinyl group, so needs to be converted to carboxyl group prior to bonding. Wet chemistry methods were found to have poor yield and poor bonding results for this surface reaction, so an ozonation process was developed [21] to attack the double bond preferentially and produce a hydrolysable surface with good yield. If too much ozonation is used, the layer is eliminated, so process control is required. A water rinse for hydrolysis completes the conversion to carboxyl termination. The reason that a SAM with carboxyl termination was not used initially is that it is difficult to produce high quality monolayers when both ends of the molecule are polar.

The covalent bond between these SAM layers is formed in the reaction of the two different functional groups, or monomers, located at the ends of the two organic SAMs molecules. The functional groups at the ends of these SAMs, a carboxylic acid and an amine, figure 1, combine through a condensation reaction, similarly to the way in which amino acids form peptide bonds with the same functional groups (it is also used in nylon production). The chances of getting covalent bonding between the layers are favorable if the functional groups can be brought into contact. In figure 2, two-inch wafers were bonded without the use of an intermediary. The ~5 cm square wafers were placed with active terminated sides together, weighted by ~125 N, and heated to 285 °C for 1 h. The 'bare' bond is achieved, but only in regions in which the wafers were touching each other, as they are thick enough not to deform sufficiently to make intimate contact over the entire surface area. The bond strength indicates a covalent connection between the wafers. This sample was pulled apart manually, using a sharp implement to pry apart the wafers. On the left of figure 2, the wafer was coated with carboxylic acid-terminated SAMs, and on the right, the wafer carried the amine-terminated SAMs. The bonding occurred in millimeter-sized regions. This is presumably the contact area given the elasticity and flatness of the wafers on the scale of the ~1 nm SAM layer thicknesses [22]. The nanoglue bonding of the SAMs exceeds the silicon-silicon bond strength, as shown by the transfer of parts of one wafer to the other in the bonded regions. This could be over-come by use of a thinner top wafer, which would deform much more easily and require less pressure to bring the surfaces together. The wafers did not bond without heat and pressure (even very thin wafers), which increases the ability to control their relative positions and verifies that the covalent bond is indeed forming. Very thin wafers are known to bond to thicker, clean wafers almost immediately. If this nanoglue was used, the bond would be as strong, but they could be positioned and bonded only when properly aligned. The fact that bonding did occur where the SAMs were in contact, and was strong, implies that the constraints on coordination (angles and positions) brought about by the attachment of the molecules with the active groups to the surface does not significantly inhibit the peptide reaction. Previous work in this system is in aqueous solution, which presents no coordination constraints.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Although the nanoglue can be used on deformable wafers, we seek a more generic nanoglue. Polished wafers tend to have a roughness ~λ/10 or less for the ~600 nm test wavelength, or about ~60 nm over rather long lateral distances. The SAM molecules are ~2 nm each, only span a small fraction of this distance without wafer deformation. Our approach uses an intermediary that is covalently bonded to bridge the gap. The choice of intermediary is governed by the bonding requirements, which we now state: the bonding should be semi-permanent to permanent, although an option of a removable bond, for example at high temperature, may be useful. The bond should have mid to high temperature tolerance so it can withstand processing and operational temperatures. The bonding cannot be instantaneous. This process should only take place after the alignment has been completed. There must be constituent compatibility in this process so as not to damage devices on the wafers. Few by-products should be produced from this process or released at operational temperature after formation. Acid by-products can cause oxidation of the chips, destroying them. Gaseous by-products can create air pockets that will expand when heated, causing cyclic stress that will reduce a device's lifetime [23]. The byproduct of peptide bond formation is a small amount of molecular water (<1 monolayer), which is a small molecule that does not seem to cause problems.

With these constraints in mind, we attempted to grow Kevlar as an intermediary. Kevlar consist of long molecular chains of poly(paraphenylene terephthalamide). There are inter-chain hydrogen bonds that made Kevlar extremely strong. It is made from dissolved para-phenylene diamine (amine on both ends) and terephthalic acid (carboxyl on both ends). The problem is finding a solvent for the terephthalic acid that does not violate the above conditions. We used 0.2 molar solutions of para-phenylene diamine in water and terephthalic acid in 5% by volume sulfuric acid in water. Bonding was observed and was occasionally very strong, but was often unsatisfactory due to the solvent problem.

The nylon system is much easier to adapt to our nanoglue. It is made from dissolved hexamethylene diamine (amine on both ends) and adipic acid (carboxyl on both ends). The former dissolves in water and the latter in ethanol, and we made 0.0295 molar solutions of each for bonding. Both remain dissolved in the mixture of these solutions. All documentation on nylon formation is for free molecules and the conventional recipe for nylon requires an autoclave, i.e. wet conditions and elevated temperature and pressure. Such conditions may be unsuitable for wafers. In particular, the water will evaporate and form gas pockets that will reduce bond lifetime. In solution, the carboxylic acid of the adipic acid donates a hydrogen to the amine of the hexamethylene diamine at room temperature. Together, they form an ionic bonded nylon salt. We found that we can make this salt in solution, then dry it out before forming the bond, alleviating this problem. To form nylon, which drives off a water molecule and covalently bonds the C to the N, the samples are heated to the polymerization temperature, ~150 °C–285 °C. The amine or carboxyl group on the SAM surfaces will participate as the others in solution, and covalently link both surfaces if the chain is continuous across the gap. Thinner layers increase the probability of this. Thicker layers will likely result in what resembles a standard glue—entangled polymer fibers, which is a temperature removable glue using the same system described here. The thicker layers were formed by evaporating the solvent from a mixture of the precursor solutions until a wet milky-white substance formed, then placement of that between the wafers that were pressed together (~15 atm equivalent) while heated to 290 °C for 30 min, until the water was driven off. The wafers reliably (non-covalently) bonded and were thermally removable, but the bond strength (0.4–2  ×  10⁵ Pa) was not well controlled.

The covalent nanoglue requires a thinner layer with more order. Equal amounts of 0.0295 M precursor solutions were combined, and ~0.05 ml dropped on to the lower wafer. The 0.5–1 cm² top wafer was lowered onto the lower wafer, and the two clamped with a small force (equivalent to an alligator clip), and left 24 h. This allowed the solvent to slowly evaporate and an ionic crystal to form. We believe that the crystal is seeded by the upper and lower wafers, so is oriented to enhance the probability of covalent connection between the wafers. If a top wafer is not used, a crystalline pattern was observed on a bottom wafer. Polymerization is induced by placement in an oven at 200 °C for 30 min. A consistent bond is produced, that either has a strength of ~180  ±  20 kPa, or the strength is larger than our cantilever force meter can measure, >  ~2 MPa. The latter are the covalent nanoglue bonded cases. The yield stress of nylon fiber is ~50 MPa, so the better samples might be covalent bonded from surface to surface. If we wedge apart the surfaces, which takes forces that would suggest nylon yield (and often results in broken silicon), regions in which the nanoglue has failed always show similar residue patterns on both sides, indicating that the bond to the wafers is intact and that the intermediary, nylon, layer has failed. This strongly supports the assertion that the wafer is covalently bonded to the SAMs and that they are covalently bonded to the intermediate material. To make the layer thinner, we spin the mixed precursor solutions onto the wafers. This cannot be done under ambient conditions, since the solution would quickly dry out and leave a thick, uneven or partially uncoated 'film.' We spin under  >85% relative humidity (using a hard disk spinner), which results in reproducible thin films that are then clamped while still in the wet box, and treated as before. The film thickness is ~40 nm for 1.8 mM precursor molecules, and ~45 nm for 7.4 mM precursors. Higher concentrations do not give a narrow distribution of film thicknesses. The process was not optimized and many samples did not bond, but those that did tended to be similar to the stronger bonds of the 'drop' method described above. The order induced by the crystallization of the salts increased the fraction of bonds too strong fo our force measurement system, which we assume to be covalent bonds wafer-wafer.

We can characterize the temperature dependence of the bond strength for those bonds that were in range of our force sensor: the non-covalent-bonded or temporary bonded wafers. A large number of samples were prepared under similar conditions. The stage of the force measurement apparatus was heated to several temperatures and measurements of the force needed to break the bond measured. The force was converted to a bond strength using the upper wafer area. The results are shown in figure 3 and a decreasing trend with increasing temperature is noted. The scatter in the plot is due to the use of several different samples at each temperature (each test results in a failure) and gives an indication of the level of reproducibility of bond strength.

Zoom In Zoom Out Reset image size

The quality of the bonding layer can be investigated by through wafer imaging with IR light and double-polished wafers. An incandescent bulb with the IR filter removed from the microscope is used as a light source. The silicon filters all light with wavelength shorter than ~1.1 microns. A Si CCD camera is used for detection, without an IR filter. Its response declines rapidly above ~1.1 microns, so the source and camera limit the investigation to a narrow wavelength range. This enables interference measurements at low magnification in a Nikon dissection microscope. The interference fringes, figure 4, are similar to Newton's rings, and are due to the variation of the thin layer of nylon between the silicon wafers. The change in thickness between two bright fringes Δd  =  λ/(2n)  =  350 nm, where the wavelength is the 1.1 microns and the index of the film (nylon) n  =  1.565. This tells us more about the flatness of the silicon rather than the bonding layer, however. Jumps in the pattern can correspond to voids or defects in the bonding layer. We see such a jump in the lower right corner of the sample/image in figure 4, by ~1/2 fringe. If this is a void, then the jump is due to an index change Δn  =  0.565 rather than a thickness change, so 2dΔn  =  λ/2 or d  =  490 nm.

Zoom In Zoom Out Reset image size

Small-scale defects in the bonding layer can be viewed with the same illumination scheme, but with a higher resolution microscope, in this case a Nikon TE-2000 inverted microscope with the IR filters removed. The 20×  objective has a focal length small compared to the our 15–50 micron thick wafers, so residue on the wafer surfaces can be differentiated from defects at the interface by focusing. We have observed point defects, at the resolution limit of the objective, <1 micron, that tend to occur along arcs, separated by several microns along the arcs. Large areas of defect-free bonding are observed between these grouped features. The process was not optimized to reduce these and their source was not identified.

We have demonstrated a covalent nanoglue bonding standard silicon wafers together. It is based upon amine-carboxyl functional group reactions, whose reactions we find are not impeded by their attachment to a surface layer. Self-assembled monolayers (PEDA and oxidized 10-UTS) attach these groups to the surfaces of the silicon. Bare bonding of these SAMs directly to each other and via intermediary (nylon, Kevlar) are shown. The intermediary fluid is shown elsewhere [20] to function as the alignment fluid in a wafer-level self-alignment system based upon pattering of these SAMs and capillary forces, permitting both alignment and bonding in the same processing step. Several methods were used to fabricate a bond with intermediate layers, some of which proved as strong as the bare bond. A solution precursor drop method and a spinning method for even thinner intermediate layers has been demonstrated and found to have covalent bonding between the intermediate and the silicon, through the SAM. Many cases result in bonds strong enough that our measurement apparatus could not break the bond. This indicates that forming an ionic crystal prior to forming the nylon-like intermediate layer solves the orientation problem of the intermediate layer.

The manuscript was written through contributions of all authors. All authors have given approval to the final version of the manuscript.

This work was supported by AFRL-WPAFB (Grant FA8650-04-2-1619).