Hard-to-Change Flux Factor on the Wave Solder Line

The structural problem The experiment was executed as a split-plot but analyzed as if it were completely randomized. Flux was held constant for a whole day, so the four preheat/speed combinations at one flux level all share that day's conditions - operator, ambient temperature, solder pot chemistry, board lot. Those day effects are perfectly confounded with flux, and the sub-plot runs within a day are not independent of each other.

Consequences of analyzing it as a CRD

• The flux effect is tested against the sub-plot (residual) error, which is far too small for a whole-plot factor. Its F-value is inflated, so p = 0.001 is optimistic and may not survive correct analysis.
• The flux x speed interaction involves a whole-plot factor, so its standard error is also understated.
• Any day-to-day drift is silently attributed to flux.

How it should have been structured

• Declare flux as a hard-to-change (whole-plot) factor, randomize the order of flux levels across days, and replicate whole plots - at least 2-3 separate days per flux level, so the flux comparison is not a single-day comparison.
• Analyze with a mixed model containing two error strata: whole-plot error for flux, sub-plot error for speed, preheat, and the interactions. Equivalently, a split-plot ANOVA with the correct denominators.
• If flux truly cannot be replicated across days, treat it as a block and accept that its main effect is not estimable - do not report it as significant.

Message for the team: randomization must match the physical constraints. When a factor cannot be randomized run-to-run, the analysis must say so.

Judgment: reject it as stated

• 0.7 m/min is outside the tested region (0.8-1.0). The fitted model is linear in speed, but there is no data to support the shape below 0.8, and the significant flux x speed interaction means the speed slope itself depends on flux. Extrapolating a linear model past the design space is one of the classic DOE failure modes.
• The predicted 12 per 1,000 is a model extrapolation, not an observation. It should be treated as a hypothesis to be tested, not a result to be implemented.
• A lower conveyor speed changes other physics: more preheat dwell, different thermal profile, possible board warp, reduced throughput, and possibly new defect modes (dross, icicles, cold joints) that this experiment did not measure.
• Economics matter. A 4-hour flux changeover means flux B should be run in campaigns. If the throughput loss from 0.8 to 0.7 m/min costs more than the bridges saved, the economic optimum is likely flux B at 0.9 or 1.0, not 0.7.

Next runs

1. Confirmation runs at the two credible settings: flux B at 0.8 and flux B at 1.0, replicated on separate days.
2. Augment the design with speed 0.7 and 0.9 (and optionally 1.1) at flux B, with fresh center points to test for curvature and lack of fit.
3. Measure the full response set, not just bridging: total defects by type, joint strength or cross-section, ionic contamination, throughput in boards per hour, and rework hours.
4. Compare cost per 1,000 joints, including changeover time, before choosing.

Only after confirmation runs land inside the acceptance window should the setting be proposed for a pilot.

Pre-defined acceptance criteria (agreed before the pilot starts)

• Bridging at or below 12 per 1,000 joints, with the 95% upper confidence bound also at or below the level implied by 96% yield - a point estimate alone is not enough.
• Total first-pass yield >= 96%, and no increase in any other defect category (cold joints, icicles, dross, contamination).
• Joint quality passes visual and cross-section acceptance; flux residue passes ionic contamination and compatibility checks with conformal coating.
• Throughput at or above the current baseline; if it drops, the business case must still hold on total cost.
• Process parameters stay inside the machine's capability window (preheat, pot temperature, wave height) across a full shift.

Pilot design

• One line, two to three production lots per setting, randomized run order, same operators and boards, defect counting blinded or verified by attribute agreement.
• Run the current setting and the proposed setting side by side rather than sequentially by week, to avoid confounding with lot or season.
• Written stop rules: any safety issue, any out-of-spec joint quality, or two consecutive lots above the bridging threshold ends the trial.

Before scale-up, also

• Update the PFMEA with the new failure modes and re-score; mistake-proof flux selection with a recipe interlock so the wrong flux cannot be loaded.
• Update the control plan with the new settings, monitoring method, and reaction plan; train operators on the changeover procedure.
• Verify flux B supply, shelf life, cost, and lead time - a yield gain that creates a single-source dependency is not a clean win.
• Scale line by line with a confirmation lot at each line, not all at once.