Reading time ( words)
Proper application of solder paste onto a PCb is essential in surface mount assemblies. this article examines a screening design of experimentation (doe) using an evolutionary operation (evop) approach to understand this process, and ultimately improve electrical yields.
By Bill Craig
Applying solder paste correctly to a printed circuit board (PCB) is one of the most critical steps in surface mount assembly. This article highlights a screening design of experiment (DOE) that was used with an evolutionary operation (EVOP) approach to gain insight into the process. To minimize electrical test failures, SMT assemblers must optimize and control solder paste volume. One important factor proved to be that how an experiment is conducted is just as important as how well it is designed. By automating the stencil cleaning method, significant improvements were achieved. Whether running an experiment or a full-scale production, removing operator-induced variation is important.
With the ultimate goal of improving electrical yields and minimizing customer returns, many experiments were conducted at the screen printing process. Because many of the defective PCBs could be fixed with a simple solder joint touch-up, improvements were needed at the screen printing operation. As a result of an in-line paste monitoring system, an association between paste volume and solder defects was discovered. Significant departures from the nominal solder “brick” resulted in either opens or shorts.
The stencil aperture design studied had a nominal solder-paste buildup volume of 2880 mils3. Experiments were consistently obtaining very low R2 values of 0.3 or less, which suggested that the experimental model did not account for much of the variation. Therefore, the cleanliness of the stencil bottom was suspected as the cause for this variation from one run to the next. Manual stencil cleaning was inconsistent, and it was difficult to decide on a cleaning frequency that would not lead to false conclusions in the DOE. At this point, converting to the automatic stencil-cleaning mode was applied. The approach was similar to EVOP and consisted of administering incremental changes to the production process instead of using an off-line engineering approach. The production process was changed so that the point at which the stencil would automatically be cleaned varied between experiments. Because there was a cost associated with the rolls of lint-free paper used for wiping the stencil bottoms clean, the interval was incrementally changed from two to three. Finally, it was changed to four PCBs between cleaning cycles while monitoring paste quality and yields. Once it was determined that the ideal automated cleaning process was an interval of four PCBs during production, the next experiment could be integrated within a daily production run.
Experimental Design
Previous experiments were four-factor, two-level factorials. To study the effect of stencil cleanliness, a fifth factor, E, was introduced into the matrix. The factors and definitions are:
- Factor A: Squeegee Speed-The rate at which the squeegee blade moves across the stencil.
- Factor B: Squeegee Force-The force differs with requisite blade length. A recommended force-over-blade-length ratio is approximately 1:5.
- Factor C: Down Stop-The distance the squeegee blade is driven beyond the stencil aperture.
- Factor D: Snap-off Distance-The separation distance of the stencil from the PCB after each print stroke.
- Factor E: Stencil Condition-Runs with this factor set at Pre-clean (-) were conducted just before stencil cleaning. Runs set at Post-clean (+) were conducted just after stencil cleaning.
With the screen printer programmed to clean the stencil bottom of every fourth PCB during production automatically, half the runs were conducted just after stencil cleaning, and half were conducted just before cleaning. The 25-1 fractional-factorial design resolved main effects and two-factor interactions, but not higher-order interactions, which are extremely rare. This type of design yields highly reliable results, but fewer runs. Thus, a single replicate of a 25 full-factorial design requires 32 runs. In comparison, the one-half, fractional-factorial design tested all five factors in 16 runs.
The base design (Table 1) shows 24 full factorials for factors A through D. By adding Factor E as the fifth variable without increasing the number of runs, the final design becomes a 25-1 fractional factorial. This 25-1 fractional factorial includes a subset of the 32 possible treatment combinations, causing some of the effects to be aliased. Due to the number of runs required, full-factorial designs are inefficient. Fractional factorials allow experimenters to study many factors with a minimum of runs, which is why the 25-1 fractional factorial is used here. The high/low levels of Factor E were determined by multiplying the signs of the first four columns.
Data Analysis
The data in Table 1 were analyzed using design software*. It should be noted that this experiment ran in randomized order (second column) at the normal production rate of 45 units/hour. The last column, or response, represents average paste volume for the 12 measurement sites on a PCB. The DOE software calculated the effects of Factors A through E and their interactions (for example AB, AC). The sign in Factor E’s column is applied to the value in the response column, then all 16 values are summed accordingly. With eight Factor-E runs at the high level, or Post-clean, and eight at the low level, or Pre-clean, the average effect was reached by dividing by eight.
The DOE software created a probability plot of the effects (Figure 1), showing Factors A and E to be significant statistically. All other main effects and interaction effects are clustered along a straight line, indicating they are not significant and are caused by normal process variation.
FIGURE 1. Normal Plot reveals the effects of Factor E and Factor A to be unusually large.
Analysis of variance (ANOVA) indicated model significance, with a calculated F-ratio of 47.17. This is well above the threshold of 3.8 for 5% risk (95% confidence). The associated probability (“p”) value was less than 0.0001, which factors far below the generally accepted risk level of 5% for statistical significance. The R2 value also was calculated at 0.87, which indicates that most of the variation in the DOE was explained by the model.
Conclusion
Stencil apertures had a theoretical volume of 2880 mils3. The prediction data from the DOE revealed that volume varies from 2808 to 3023 between stencil cleaning cycles when squeegee speed (Factor A) was set at 1; and Factors B, C and D were held at their nominal levels:
- Factor B: 12
- Factor C: 75
- Factor D: -5.25
Conducting automated stencil cleaning after the fourth PCB essentially reset the process to 2800 mils³. Also, avoiding the extreme values of paste volume (below 2000 and above 3800) helped minimize solder defects.
Automated stencil cleaning proved to be beneficial. Production increased by 25% to approximately 56 units/hour, while demonstrating reductions in solder opens and shorts. Maintaining stencil-bottom cleanliness is necessary for six sigma quality. In addition, changing the process from manual cleaning to automated cleaning improved solder-paste volume control, which immediately impacted the defects per million opportunities (DPMO).
*Design-Export software, Stat-Ease, Inc.
Bill Craig, Global Front-End Six Sigma Quality Manager for ON Semiconductor, may be contacted at bcraig6sig@comcast.net.
FIGURE 1. Normal Plot reveals the effects of Factor E and Factor A to be unusually large.