1

Assessment of XRF Technique as a Method to Measure Percent Ag in SnAg Solders for Flip Chip Applications

Jennifer D Schulera Chia-Hsin Shihb, Charles L Arvina, KyungMoon Kimc, Eric Perfectoa

a Semiconductor Research and Development Center, IBM, Hopewell Junction, New York12533

b STATS ChipPAC Taiwan Ltd, Hsin-Chu Hsien Taiwan, R.O.C 307c STATS ChipPAC Korea Ltd, Kyoungki-do, 467-814 South Korea

Pb-free SnAg solder has become the industry standard for fabricating flip chip interconnects utilizing C4 (controlled collapse chip connection) technology. One area of interest for manufacturability of Pb-free solders is the ability to control and measure the %Ag composition and its variation from wafer to wafer, chip to chip, and C4 to C4.

There are various ways to measure solder composition. These are divided into two categories which are invasive and non-invasive referring to whether solder must be removed from the wafer in order to conduct the measurement.

There are a variety of invasive methods including Atomic Absorption (AA), Differential Scanning Calorimetry (DSC), Inductively Coupled Plasma (ICP) and Electron Probe Micro-Analyzer (EPMA) used with cross sections. Non-invasive methods are limited, making the development of the non-invasive X-Ray Fluorescence (XRF) method an important technique to determine both the thickness and composition of C4s on wafers without modifying the wafer.

There are many factors which can affect the accuracy of the XRF measurements. These include bump geometry, composition, UBM(under bump metallurgy) stack, bump spatial density, underlying chip wiring, tool vibration and tool parameters, such as collimator size, power levels, scan time, etc.

This paper will address the implementation issues in utilizing XRF for Pb-free solder SnAg systems. The paper will describe:

(1) Experimental bumping variables, (2) XRF configuration, calibration, optimized measuring

methodology and the importance of having known standards with the same dimensions of the bumps being measured

(3) Measuring accuracy and correlation with ICP and DSC, (4) Ag distribution study in the die and wafer level

2

Sample Preparation Several experimental bumping variables impact the XRF measurement, including underlying metallurgy, BLM (bottom layer metallurgy) size, UBM stack, C4 height and Ag% target (Table 1). To fully understand the various factors, two kinds of wafers wereused. One was a silicon oxide wafer which was classified as mechanically good (MG). The other type was with underlying chip wiring and front-end processing, which was classified as electrically good (EG). The test wafers received a sputtered seed layer, and then received a plated UBM and SnAg solder. To understand the effect of the UBM on the XRF reading, two UBM stacks were prepared containing a thick Cu UBM or no Cu UBM.

Vehicle No.

Wafer Type BLM(µm)

C4 Height(µm)

UBM Stack %Ag Target in SnAg Solder

1 MG 90 70 No Cu SnAg (0.6%Ag)

2 MG 90 70 No Cu SnAg (1.7%Ag)

3 MG 90 70 Thick Cu SnAg (0.6%Ag)

4 MG 90 70 Thick Cu SnAg (1.7%Ag)

5 EG 90 70 No Cu SnAg (0.6%Ag)

6 EG 90 70 No Cu SnAg (1.7%Ag)

7 EG 90 70 Thick Cu SnAg (0.6%Ag)

8 EG 90 70 Thick Cu SnAg (1.7%Ag)

9 MG 110 90 Thick Cu SnAg (0.6%Ag)

10 MG 110 90 Thick Cu SnAg (1.7%Ag)

11 EG 110 90 Thick Cu SnAg (0.6%Ag)

12 EG 110 90 Thick Cu SnAg (1.7%Ag)

Table 1. Test Vehicle List. MG: Mechanically Good. EG: Electrically Good.

In order to setup and calibrate the XRF settings, C4NP standards were constructed[2]. C4NP is a solder transfer process which controls the Ag% composition to ± 0.1%Ag. ICP was used to confirm the samples produced by C4NP (Table 2).

No.BLM Size

C4 Height

Wafer Type

Known Bump Composition

UBM Stack

SnAg Solder Process

A 90 70 MG 0.5%Ag Ni/SnAg C4NPB 90 70 MG 1.8%Ag Ni/SnAg C4NPC 110 90 MG 0.5%Ag Ni/SnAg C4NPD 110 90 MG 1.8%Ag Ni/SnAg C4NP

Table 2. Standards For XRF setup & calibration.

Tool ConfigurationThree tools were utilized to determine bump composition (Table 3). These were

the XRF, ICP-OES (optical emission spectroscopy) and DSC. Through data comparison with ICP and DSC, a non-invasive XRF method was created and calibrated. At the same time, procedures were created to allow for daily checks and future validation and re-calibration of the XRF unit if readings began to drift.

3

Tool Description

XRFAnode material is made of tungsten or molybdenum

ICP-OESInvasive method to get C4 composition on global zone

DSCInvasive method to get C4 composition on global zone

Table 3. Bump Composition Inspection Tool Information.

It is important to understand the general workings of the XRF unit in order to write the XRF recipe. The X-ray tube generates the primary X-radiation (primary radiation). The electrically heated cathode emits electrons. Accelerated by the applied high voltage to very high speeds, the electrons bombard the anode material. This generates the primary X-radiation. The primary filter optimizes the energy distribution of the primary X-radiation. The shutter serves as a safety device and closes the access of the primary X-radiation to the measurement chamber, if needed. Within the recipe, it requires one to pick the type of filter, the accelerating voltage, the power, etc. Each one of these values were chosen based upon the type of structure to be analyzed. It was extremely important to know how each impacted the accuracy of the final reading. In many cases, it was unknown how each new structure would react. Thus, much trial and error was required initially until the final recipe could be determined. In all cases, it was paramount that standards of the correct size and geometry were utilized.

XRF Recipe SettingThe Si PIN Diode detector with the advanced de-convolution software makes it

possible to reduce the number of standards required. However, even with this advance, it was important to establish the recipe correctly and have at least two standards of the correct geometry and structure. The need for the correct geometry comes from the learning that the %Ag is mass transport controlled. Thus, the smaller the opening in the photoresist, the %Ag in the final deposit is reduced. In addition, for a constant plating rate, as the plating gets closer to the top of the photoresist, the %Ag in the deposited SnAg increases. This means that the concentration of Ag in a test structure on the edgeof a wafer may have a different and most likely higher concentration of Ag than the Ag in an actual C4.

Establishing the recipe first required defining the structure to be measured. This

included listing the various stacks of layers if they could contribute to the intensity of a given element and if the layer was a single element or a combination. Initial calibration was conducted in accordance with the recommended procedure in the tool manual using the C4NP standards. Based on experimental data, the initial recipe process time was set to180s (Figure 1). Unstable readings were detected on the C4NP standards A and B with 70µm C4 height (Table 4). As a result, standards C and D were used for calibration and further investigations were conducted to determine the root cause of the unstable readings on standards A and B.

Tool issues such as stage movement accuracy and laser alignment were investigated and then eliminated as possible causes. The next item looked at was the impact of the sample itself. The X-ray spot size of 40µm was an appropriate dimensionto measure standards A and B. Thus, a C4 geometry issue was hypothesized. To prove

4

this, the C4s from standards A and B were stamped to make them flat on the top surface to alleviate X-ray scattering effects. Reliability dramatically improved after the flattening, proving that C4 geometry impacts the intensity of the signal. (Table 5, Figure 2, Figure 3).

Figure 1. Longer recipe process time creates a stronger XRF feedback generating more stable readings. The condition with recipe process time 1000 sec is best. However, 180 sec produces a comparable standard deviation and significantly reduces the data collection time from 17 min to 3 min.

Process time 180sec on C4NP Standards 20 XRF Readings on Randomly C4sNo. BLM Size C4 Height Ag% Target Mean Std Dev RangeA 90 70 0.5% 0.66% 0.44% 1.84%B 90 70 1.8% 1.77% 0.16% 0.53%C 110 90 0.5% 0.48% 0.06% 0.19%D 110 90 1.8% 1.79% 0.05% 0.21%

Table 4. XRF inspection on C4NP Standards

Process time 180sec on flattened C4NP standards

20 XRF Readings on Randomly C4s

No. BLM Size C4 Height Ag% Target Mean Std Dev RangeA 90 70 0.5% 0.50% 0.05% 0.20%B 90 70 1.8% 1.89% 0.06% 0.27%

Table 5. XRF inspection on flattened C4NP Standards

The spot size of the X-ray opening is 40 µm. However, this is not the same as for the detector. Thus, the detector is collecting photoelectrons from any region that is excited by the X-rays. Thus, by stamping, it reduces the chance of exciting larger regions which would cause artificially higher readings and thus a larger standard deviation. This

5

concept is further elaborated in the next few sections by also taking into account the background noise.

Un-stamped Flattened

Figure 2. The SEM photos for the comparison between un-stamped and flattened C4s.

Figure 3. The spectrum data shows a stronger feedback of X-ray fluorescence radiation on larger C4s.

6

The XRF region of interest was restricted to read the X-ray fluorescence range of Kα, Kβof Sn and Ag only in order to more accurately measure alternate UBM stacks since there is a high risk of spectrum overlay among Lα, Lβof Sn, Ag and other materials in the UBM stack. Voltage and current must also be balanced to keep the linear range on the detector since the voltage was needed to cause the electron transitions and current was directly proportional to intensity. If at all possible, it is best to choose conditions that can cover all analysis instead of changing them since changing tube conditions affects overall X-ray stability which in turn would affect each calibration. Voltage also affected the depth of penetration of X-rays. With most XRF spectrometers, X-rays penetrate deep into the sample to generate X-ray fluorescence. The fluorescent X-rays coming out of a sample depend on the energy of the incoming X-ray and the travel distance to the detector. Often X-rays are self-absorbed within the samples and never reach the detector.As energy increases, the penetration is deeper into the sample. However, there is a point known as infinite thickness for each element where the photoelectron intensity leaving the sample does not change with thickness. This is considered a bulk type analysis. High power at 50kV was used to excite the electrons on K layer to enhance the peaks of Kα, βof Sn and Ag (Figure 4).

Figure 4. Generation of X-ray fluorescence radiation.

Since flattened C4 tops provide a stronger XRF signal feedback, measurements were attempted prior to reflow. Unfortunately, the diffusion limited plating properties of Ag% were highlighted in this study. Ag deposition at the initialization of the stack is haphazard, but gradually the Ag% increases within the deposit from bottom to top. The highest Ag% was found on the top of the C4 column. The XRF penetration may not be deep enough to reach the bottom of solder column. This generated inconsistently high Ag% readings relative to homogeneous reflowed bumps.

Results

Data InterpretationExamining the XRF data (Table 6), the flattened C4 top method was necessary to get

stable and exact readings on samples with low C4 height <=70 µm. XRF readings on MG wafers are comparable with ICP data (within 0.2%Ag), but the XRF readings are not reliable on EG wafers. According to the fluorescence spectrum, high background noise is found on some C4 positions with underlying chip wiring on EG wafers. Rebuilt C4NP EG chips confirmed the background noise issue. Underlying chip wiring caused high background noise in XRF spectrum (Figure 5, Figure6). To alleviate this, the XRF collection time was shortened to mitigate primary X-ray penetration (Table 7). If the

7

SnAg thickness was sufficient (in this study, 90 µm), good reliability was found on XRF readings for both 0.5% and 1.8% Ag.

Wafer Type

BLM(µm)

BH(µm)

UBM Stack(µm)

%Ag Target

Unstampedor Stamped

XRF Reading (%) ICP

(%)Mean σ

1 MG 90 70 No Cu Plated0.6%Ag

Unstamped 0.78 0.450.63

Stamped 0.66 0.05

2 MG 90 70 No Cu Plated1.7%Ag

Unstamped 1.83 0.221.72

Stamped 1.71 0.06

3 MG 90 70 Thick Cu Plated0.6%Ag

Unstamped 0.84 0.240.73

Stamped 0.74 0.03

4 MG 90 70 Thick CuPlated1.7%Ag

Unstamped 1.93 0.321.71

Stamped 1.73 0.05

5 EG 90 70 No Cu Plated0.6%Ag

Unstamped 1.04 0.370.65

Stamped 0.80 0.22

6 EG 90 70 No Cu Plated1.7%Ag

Unstamped 2.23 0.611.77

Stamped 1.97 0.32

7 EG 90 70 Thick Cu Plated0.6%Ag

Unstamped 1.22 0.560.76

Stamped 0.97 0.19

8 EG 90 70 Thick Cu Plated1.7%Ag

Unstamped 2.33 0.711.65

Stamped 1.96 0.16

9 MG 110 90 Thick Cu Plated0.6%Ag

Unstamped 0.44 0.040.46

Stamped 0.45 0.05

10 MG 110 90 Thick Cu Plated1.7%Ag

Unstamped 1.56 0.04 1.52Stamped 1.52 0.04

11 EG 110 90 Thick Cu Plated0.6%Ag

Unstamped 2.11 0.700.41

Stamped 0.63 0.30

12 EG 110 90 Thick CuPlated1.7%Ag

Unstamped 2.06 0.401.50

Stamped 1.60 0.20

Table 6. Experiments. Sample size: 20 random C4s/1 chip/Test Vehicle on plated chips.

Wafer Type

BLM(µm)

BH(µm)

UBM Stack(µm)

%Ag Target

Process Time

XRF Reading (%)

Mean σRange

13 EG 90 70 2Ni C4NP 0.5%Ag

30 sec 0.66 0.32 1.10180 sec 1.29 0.42 0.99

14 EG 90 70 2Ni C4NP1.8%Ag

30 sec 2.23 0.24 1.05180 sec 2.00 0.12 0.90

15 EG 110 90 2Ni Plated0.5%Ag

30 sec 0.53 0.07 0.24180 sec 0.69 0.28 0.92

16 EG 110 90 2Ni Plated1.8%Ag

30 sec 1.90 0.06 0.24180 sec 1.90 0.06 0.17

Table 7. Experiments. Sample size: 20 random C4s/1 chip/Test Vehicle on C4NP EG chips.

8

Figure 5. High background spectrum noises are found on EG chips.

Figure 6. Illustration explaining spectrum performance between EG and MG chips and why high Ag% is detected on EG chips.

Prior to flattening C4 Post flattening C4

Figure 7. Flattened C4 can help obtain a stronger XRF peak. The spectrum post flattening is better than an unstamped C4

EG, 70 µm C4, 0.5%Ag, C4NP

MG, 70 µm C4, 0.5%Ag, C4NP

ROI 644~920

9

Dissolved Cu% in SnAg SolderIn order to determine how much Cu is dissolved into the bulk solder, two techniques,

DSC and XRF are used. In order to obtain the melting point of the bulk solder, it must be separated above the Cu intermetallics (IMC). DSC was then conducted to obtain the solidus temperature, liquidus temperature along with the melting and solidificationbehaviors. By knowing the liquidus temperature and the Ag% obtained from XRF, aSnAgCu phase diagram could then be used to obtain the Cu%. (Figure 8).

The phase diagram in Figure 8 is an equilibrium diagram. However, DSC is collected at various rates. Three different scanning rates were used to determine the impact of the melting point for each alloy as a function of scan rates shown in Figure 9. Figure 10 shows how to relate the DSC scan to the phase diagram. As was expected, the solidus temperature was independent of the scanning rate. However, the liquidus temperature was dependent. In order to use the phase diagram, one needs to determinethe melting point at equilibrium. It can be seen in Figure 10 that the temperature of the final melt is reduced as the scan rate is lowered. The reason for this is not allowing enough time at each temperature step for equilibrium to occur. However, the step change in temperature as a function of scan rate is being reduced. The tool available for this work could only go down to 5oC / sec. Determining the final melt temperature and each of the scan rates of 20, 10 and 5 oC / min, it was possible to obtain the following relationship for this particular structure, y = 0.1422 x + 229.75. Utilizing this relationship and extrapolating to near zero showed that only a further 0.7 C reduction should occur from the measured value at 5oC / sec. As such, all samples were run at 5oC / sec and 0.7 degree C was subtracted from the measured value. It was also observed that there are two peaks in Figure 8 and 9g. This was due to the various phases within the alloy. The first peak is the transformation of the Ag3Sn / beta Sn interface from solid to liquid which is the eutectic composition and melts first. The second peak is the transformation of the remaining material. As the %Ag increases, the relative values of the first peak to second peak change. In the case of the 1.7% Ag alloys, the second peak is just a plateau.

10

Figure 8. Based on the DSC data and ternary phase diagram, the rough dissolved Cu% range in bulk solder could be determined

Figure 9. The DSC curves under different heating rates. (5,10,20 degree C/min)

11

Conclusions

XRF is a popular, non-invasive inspection method for bump composition. Several challenges were encountered when establishing this technique. A key item needed to implement this method was how to balance X-ray power voltage and current settings to obtain suitable X-ray penetrating ability. If the X-ray penetration was too high, background noise was intensified in the presence of underlying wiring of EG chips. If the penetration was too low, unstable readings were encountered and there was low accuracy. There existed a permanent background noise baseline creating a tool ability limitation for very low Ag% detection (less than 0.2% Ag). Through XRF, ICP, DSC comparison, XRF recipes can be developed as an excellent monitor for non-invasive bump composition evaluations. As in all methods, XRF has a unique set of challenges and limitations which can be overcome through proper calibration and verification.

As the packaging trend moves toward finer pitched products, studies of small diameter C4s will become pervasive. If XRF is preferred to measure the bump composition of smaller C4s, geometry will become a key issue. C4 flattening helps mitigate geometry issues on smaller C4s, but may be insufficient for micro-bump features. Future work will involve how to define C4 geometry and the possible creation of composition test sites which will need to have extensive correlation established.

References

1. J. Sylvestre, et al.,”The Impact of Process Parameters on the Fracture of Brittle Structures During Chip Joining on Organic Laminates,” 2008 ECTC.

2. E. Perfecto, et. al, “C4NP Technology: Present and Future,” 2008 IMAPS Device Packaging Conference.

3. B.. Beckhoff B. Kanngießer N. Langhoff R.Wedell H.Wolff (Eds.) in Handbook of Practical X-Ray Fluorescence Analysis 2006.

Liquidus Line

Solidus

0.6% Ag

1.7% Ag

Figure 10. By rotating the DSC curves 90o it is possible to see how they relate to the phase diagram. Note that as the Ag% increases, the size of the first peak increases and the size of the second reduces.

BA