Chip Package Interaction: Understanding of Contributing Factors in Back End of Line (BEoL) Silicon, Cu Pillar Design and Applied Process Improvements

The initiation of cracks in the brittle ultra low-k dielectric material in the Back End of Line (BEoL) on advanced node silicon devices is of major concern for assembly processes [1–4]. It is attributed mainly to various combinations of the Chip-Package-Interaction (CPI) effects. This challenge is further amplified by the adoption of Cu Pillars to replace conventional solder bump flip chip interconnects as the device bump pitch shrinks and the demand for higher I/O counts per area soars. Furthermore, the adoption of Cu Pillars as interconnects is inevitable because Cu Pillars offer superior electrical performance. The high modulus and yield strength of Cu can transfer significantly higher thermo-mechanical stress to the ultra low k layer and increases the risk of dielectric cracks. It is therefore important to understand the CPI challenges of Cu Pillars on ultra low k chips in detail, and more importantly to turn all possible knobs to mitigate the risk of package induced failure modes in the BEoL.

Fig. 1 depicts the main contributing CPI risk factors, which are related to the silicon (i.e. BEoL stack and design rules, processing of the silicon) the interconnect material and design rules (i.e. SnAg bumps, Cu Pillar), the assembly technology (FC-CSP, FC-BGA, wafer level packaging), the assembly material and the substrate or board design.

The broad spectrum of impacting factors is depicted in Fig. 1 and had been described in detail previously [5].

This paper concentrates on the impact of Cu Pillar design and Cu Pillar bumping process improvements as well as the contribution of BEoL process optimization, which is summarized under the terms “Interconnect Material”, “Interconnect Design” and “BEoL Material” in Fig.1.

A multi-level package modeling technique is used to verify the experimental data with respect to under bump metallization (UBM) diameter and Cu Pillar height variations. This modeling technique is already well described in a previous publication [6]. It is to be noted that a chip package interaction (CPI) failure that is usually observed in BEoL occurs within the ultra low-k (ULK) levels under the bump, which is where the maximum ULK strain is observed in the simulation model as depicted in Fig.2.

For the first case study the UBM diameter is varied between 46μm to 50μm as seen in Fig.3. The strain in the ULK level decreases consistently as the UBM diameter increases. The ULK strain is reduced by 5% and 8% when the UBM diameter increases from 46μm to 48 and 50μm respectively.

The second case studies the effect of Cu Pillar height variation from 30μm to 35μm on the maximum strain generated in the ULK. The maximum principal strain in ULK increases by 3% with 35μm Cu Pillar height as compared to 30μm Cu Pillar as depicted in Fig.4. In addition, Fig.5 shows that the area of the maximum ULK strain is increased with 35μm Cu Pillar height as compared to 30μm Pillar.

The single bump shear modeling technique, known as bump assisted back end of line stability indentation (BABSI) simulation, is used to check the effect of the UBM undercut on under bump BEoL stability with different UBM diameters. This modeling technique employed was also described previously [7].

In Fig. 6, the UBM undercut has been varied between 0 and 2μm for cases with UBM diameters varying between 40 to 80μm. Here again it is seen that the larger UBM diameter causes lower ULK strain irrespective of the UBM undercut. Furthermore, larger UBM undercuts consistently lead to higher strain in the ULK for all considered UBM diameters.

The above results indicate that the effect of the UBM undercut is the lowest for a large UBM diameter of 80μm. The effect of the UBM undercut increases substantially with smaller UBM diameters, as long as the UBM diameter is significantly larger than the polyimide opening. As shown in Fig. 6, the effect of the UBM undercut is less for the 40μm UBM diameter than for the 60μm UBM diameter. This is because the UBM landing on polyimide with 40μm UBM diameter is very close to the polyimide opening of 30μm, which is considered in the model. Therefore, the shear load coming from the Cu Pillar is not so strongly absorbed by the small UBM area over the polyimide, but is transferred through the polyimide opening onto the aluminum pad and from there into the BEoL stack. Hence the ULK strain for the 40μm UBM diameter is significantly higher than for the 60μm UBM diameter and the UBM undercut effect for the 40μm UBM diameter is slightly reduced as compared to the 60μm UBM diameter.

Moreover comparing the 80μm UBM diameter and 2μm UBM undercut to a 76μm UBM diameter and 0μm UBM undercut shows, that the UBM undercut has the same influence on the strain generated in the ULK as a smaller UBM diameter.

As shown by simulation, the UBM diameter has the strongest impact compared to the Cu Pillar height. The diameter is considered to be effectively changed by the undercut. However, the solder cap height and the solder joint formation which is not considered within the shear test and shear test simulation, contribute significantly to the stresses which can be transferred into the BEoL.

The size of the UBM diameter determines the size of the contact area between the bump and the BEoL stack. The mechanical force that is exerted at the bump is distributed across the UBM area and is creating mechanical stress in the BEoL layers. The larger this area is, the better the mechanical stress can be distributed. Hence, increasing the UBM diameter will reduce the stress in the BEoL layers and therefore reduce the risk of ULK cracking (see modelling results in Fig.6).

Another very important factor is the bump stack itself. A typical bump Cu Pillar stack consists of three layers: Cu, Ni, SnAg. Typically, the thickness of the Ni layer is not varied very much because it functions as diffusion barrier and requires a certain thickness which is independent of the heights of the Cu and the SnAg cap.

The Cu height however, can be varied. The Cu height should be as small as possible to reduce stress coming from the substrate into the BEoL. The higher the Cu is, the stronger the torque will be that the mechanical forces induce at the interface of the Cu Pillar to the BEoL (Fig.8, see also modelling results in Fig. 5).

On the other hand, a minimum Cu height is needed to ensure the required distance between the silicon die and the substrate in the package is met. A reduction of this stand-off distance is limited by the ability of the molding or underfill processes to fill the gap properly.

The size of the SnAg cap can also be used to reduce the stress in the ULK layers. The more solder volume can be brought into the gap between substrate landing pads and the Cu Pillar the more strain it can absorb. However, there are limitations, which must be taken into consideration. Very important is, that the solder is kept between Cu Pillar and landing pad and that it is not completely transformed into an intermetallic compound (IMC). If the solder is wetting the Pillar sidewall, higher stress will be induced on the silicon and the effective volume of the solder, which can absorb stress by plastic strain, is reduced. Hence, a balance needs to be found for the SnAg cap height.

Modelling results determined the directions of bump process adjustments, which were later successfully implemented in production.

The final UBM diameter can be influenced by several process parameters in the bump flow. The most obvious is of course the size of the bump diameter, determined by the reticle. Furthermore, bump lithography parameters have to be evaluated. By choosing different focus and dose settings at bump exposure, the via size and shape of the bump resist can be adjusted within certain limits. The definition of the via in the resist has to be considered as it determines the geometry and shape of the Cu Pillar. In addition, the diameter of the UBM is reduced by the following UBM etch processes. Fig. 9 depicts major bump process, which can be adjusted to minimize CPI related failures in BEoL.

The standard UBM stack consists of Ti and Cu. When the UBM is etched, especially in the Cu etch step the Cu Pillar will also be attacked. The Cu loss can be significantly greater than the actual Cu thickness in the UBM layer. A balance in the process has to be found between the complete removal of the UBM and limiting the Cu loss.

The Ti etch step is another important factor and the current process uses a standard chemistry, which causes a larger undercut. Initially, a standard chemistry combined with an additional anisotropic etch step was used to remove the Ti. Even though the Ti undercut was reduced significantly, the overall results were not sufficient. Therefore, a new Ti etch process was developed with a different chemistry to further reduce the undercut. The improvements made are depicted in Fig. 10.

The final criteria for process changes based on theoretical assumptions and wafer level testing are the assembly results for actual packages and products which are presented in the following section.

The first example describes the optimization of bump lithography settings. A split lot was executed and the modelling results were verified. The process change increased the UBM diameter by less than 5 %. At the assembly level, this change leads to a significant reduction of yield loss (Fig. 11).

Another split has been performed to check the effect of a reduced Cu height. A significant improvement in yield loss had been observed and the prediction made by modeling had been verified.

Finally, an optimized process with respect to bump lithography settings and bump stack has been rolled out to production. Yield losses, which were attributed to ULK cracking at different assembly sites, were reduced drastically as seen in Fig. 13.

Ultralow-k (ULK) and low-k interlayer dielectric (ILD) films are the mechanically weakest materials in BEoL layer stacks. Extensive engineering work is required to strengthen the interfaces between these ULK or low-k films and the adjacent passivation layers. At the end, both cohesive and adhesive strength must be sufficient so that the brittle films withstand CPI stresses acting on the BEoL layer stack without initiating flaws or cracks. The choice of suitable ILD precursors as well as UV curing predominantly influence the overall strength of the brittle ILD films, keeping the tradeoff between lowest possible dielectric permittivity and highest possible mechanical strength in mind. On a real product wafer with a multi-level Cu/low-k/ULK BEoL layer stack, further parameters like local metal densities and design related influences come into play. In this context, local mechanical probing methods are used to experimentally assess the strength of structured multi-level Cu/low-k/ULK BEoL layer stacks on wafer level. The respective BABSI test and Single-Pillar Shear Test (SPST) techniques have already been described elsewhere [8]–[9]. BABSI and SPST are applied to individual die contacts, which currently consist of Cu Pillars. Depending on the desired loading condition, a pyramidal diamond probe is brought into contact with a single Cu Pillar at its top surface (BABSI test), or a wedge shaped probe approaches the sidewall of the pillar (SPST), followed by shear displacement at a predefined and feedback controlled height above the surface of the die. One unique feature of these techniques is that they can be applied to virtually all pillar positions on a silicon die, especially on critical areas like die corners where the highest CPI stresses are suspected. The failure mode obtained after shear testing on an individual die contact is classified by means of the shape of the respective load-displacement curve F_x(X). A crack propagating inside the BEoL layer stack leads to an abrupt drop ΔF_x of the lateral force at a well-defined lateral displacement X = X_crit. This is due to the brittle fracture of low-k or ULK films inside the BEoL stack. On the other hand, if the BEoL layer stack is stable all over the shear test until X = X_max is reached, the F_x(X) curve looks smooth. In this case, there is only plastic deformation of the Cu Pillar and the Al pad, or the pillar detaches from the contact pad. Such fails are not classified as BEoL fails here. They are called far BEoL (fBEoL) fails to differentiate them from cracks inside the BEoL stack. No critical quantities are reported for fBEoL failure modes since there are no well defined abrupt events in the F_x(X) curves. The BEoL stack at test sites with fBEoL failure modes is classified as being more stable than any site with brittle BEoL crack events, i.e., fBEoL fails are treated like “no BEoL fail”. Failure modes were additionally verified by optical inspection. The number of BEoL cracks identified in this way per total number of tests results is the BEoL fail rate for a certain wafer for the chosen test condition.

The BEoL fail rates for three different wafers obtained from SPSTs are shown in Fig. 14. Several dies from the wafer center, wafer edge and half the distance between center and edge were investigated on each wafer using a height of Z = 24μm of the probe apex above the surface of the die. The average height of the Cu Pillars was h ≈ 50μm. For SPSTs, the same locations of the die contacts were chosen for all samples. The die contacts consist of oblong (oval with a flat long side) Cu Pillars with their orientation of the long pillar axis being adapted to the location on the die. SPSTs were conducted at 0deg, 45deg and 90deg shear direction relative to the long pillar axis to investigate the influence of the direction of the applied in-plane stress on the BEoL stability beneath the die contacts.

The first measurement series for each wafer in Fig. 14 contains all data for 0deg, the second for 45deg and the third for 90deg shear direction. The main outcome of this study is an overall BEoL fail rate, which is higher on wafers 1 and 2 as compared to wafer 3. Wafers 1 and 2 on the one hand and wafer 3 on the other hand are different with respect to the UV curing process within the BEoL stack (UV1 vs. UV2). The difference between UV1 and UV2 is the position in BEoL processing, when the UV cure process is carried out. All other fabrication parameters remained the same. This leads to the conclusion that the UV2 process makes the BEoL layer stack stronger than UV1. Consequently, wafers with UV2 will be mechanically more stable with respect to CPI stress.

Critical strain energy values were quantified by computing the integral ∫F_xdx between X = 0 (start of lateral motion of the shear test) and X = X_crit (at the first critical event, e.g. force jump, on the F_x(X) curve). The results depend on the shear direction and on the UV cure process as indicated in Fig. 15. The highest critical strain energy for BEoL failures on wafer 1 is detected for 0deg shear, followed by 45deg and 90deg, respectively. The same trend is observed for wafer 2, which was processed in the same way as wafer 1. This demonstrates the repeatability of the method. A comparison of the mean critical strain energies for a certain shear direction reveals slightly lower values for wafer 2 compared to wafer 1. This is consistent with slightly higher BEoL fail rates on wafer 2, probably due to a slightly higher defectivity.

Most interesting, the critical strain energy increases further for wafer 3 with UV2 process, as shown by a very few remaining BEoL crack events for 0deg. shear. All other tests on wafer 3 lead to fBEoL failure modes without any cracks in the BEoL stack, especially for 45deg. and 90deg. shear. That means, the BEoL layer stack of wafer 3 is mechanically stable under the conditions of the applied loading mode of the SPST, and the applied forces just lead to plastic deformation of the Cu Pillars and the aluminum pad or peeling off of the Cu Pillars from the pad.

Finally, dies with the improved BEoL process UV2 had been assembled and tested. A significant yield loss reduction had been observed as predicted by the shear test results on wafer level (Fig.16).

Modeling was used to predict how a change in the Cu Pillar geometry impacts the CPI failure rate. The modeling results are in close agreement with the experimental data, indicating that the models used here are well suited to predict process improvements to reduce the number of CPI related failures. The modeling and shear test results also predicted the impact of adjusting processing parameters in BEoL, such as UV correctly. The majority of process improvements described in the paper is implemented out to volume production.

The authors would like to thank the engineers in bump and test facility in FAB1 for executing the experiments for an improved Cu Pillar geometry and to reduce the UBM undercut.