What DOE is
Design of Experiments is a structured method for understanding how multiple variables affect an outcome at the same time, instead of changing one factor per test run and holding everything else fixed. Changing one variable at a time feels intuitive and is how most people naturally troubleshoot, but it's slow, and more importantly, it can't detect interactions, cases where the effect of one factor depends on the setting of another. A well-designed experiment changes several factors together in a planned, structured pattern across a manageable number of runs, then uses the results to show which factors actually matter, by how much, and whether any of them interact.
DOE is most valuable exactly where intuition is weakest: processes with several variables that plausibly interact, and a real cost to getting a production setting wrong. Reflow soldering, adhesive cure, plasma cleaning, and any multi-parameter process are common candidates in electronics manufacturing.
Core terms
- Factor: a variable being deliberately changed in the experiment, such as reflow peak temperature or conveyor speed.
- Level: a specific setting of a factor being tested, typically a "low" and "high" value bracketing the range of interest.
- Response: the measured outcome, such as solder void percentage or joint shear strength.
- Interaction: when the effect of one factor on the response depends on the level of another factor, which a one-at-a-time approach cannot detect at all.
- Full factorial: a design that tests every combination of every factor and level. Thorough, but the number of runs grows quickly with more factors.
- Fractional factorial: a deliberately chosen subset of the full factorial that still estimates the main effects and often the most important interactions, at a fraction of the run count.
The DOE process
- Define the objective and responseState exactly what's being optimized or understood, in measurable terms, and how it will be measured. "Improve reflow quality" isn't a usable objective; "minimize solder void percentage on the QFN thermal pad" is.
- Select factors and levelsChoose the variables most likely to matter, based on process knowledge, and set realistic low and high levels for each, wide enough to see an effect but within the process's actual safe operating range.
- Choose a designA full factorial for a small number of factors, or a fractional factorial or screening design when there are many candidate factors and the goal is first to identify which few actually matter.
- Randomize and runRun the experiment in random order where practical, not in the sequence that's most convenient, to avoid confusing a real factor effect with a time-based drift in the process or environment.
- Analyze the resultsIdentify which factors and interactions are statistically significant, and by how much, rather than relying on a visual read of the data alone.
- ConfirmRun the predicted optimal setting as a confirmation run before locking it into the process. A model that fits the experimental data well should also predict a new run correctly; if it doesn't, something about the process isn't fully captured.
Worked example: reflow soldering void reduction
A QFN package is showing higher-than-target solder voiding under the thermal pad. Three factors are suspected: peak reflow temperature, time above liquidus, and stencil aperture reduction on the thermal pad. A 2-level, 3-factor full factorial (8 runs) is set up:
| Run | Peak temp | Time above liquidus | Aperture reduction | Void % (response) |
|---|
| 1 | Low | Low | Low | 18% |
| 2 | High | Low | Low | 15% |
| 3 | Low | High | Low | 12% |
| 4 | High | High | Low | 9% |
| 5 | Low | Low | High | 10% |
| 6 | High | Low | High | 8% |
| 7 | Low | High | High | 5% |
| 8 | High | High | High | 3% |
Reading across these eight runs, all three factors are pushing void percentage down as they move from low to high, but the biggest single jump comes between run 1 and run 8, the combination of every factor at its high level, not from any one factor alone. That's the interaction DOE is built to catch: if only peak temperature had been tested one at a time, at whatever the current aperture and time settings happened to be, the full benefit of changing all three together would never have shown up. The confirmation run at the predicted optimum, high temperature, high time above liquidus, high aperture reduction, comes next, to verify the 3% result holds before it's written into the production profile.
Common mistakes
- Testing too many factors at once without a screening step. A full factorial on six or seven factors quickly becomes an impractical number of runs; screen down to the vital few first.
- Ignoring interactions. Treating factor effects as independent, when a one-at-a-time approach can't tell you they aren't, is the single most common reason a "confirmed" optimal setting doesn't hold up in production.
- Not randomizing run order. A process that drifts with time, oven warm-up, tool wear, can masquerade as a factor effect if every "low" setting happens to run first.
- Skipping the confirmation run. A model fit to the experimental data should be checked against a fresh run before it's trusted for production, not assumed correct because the math worked out.
Frequently asked questions
When is DOE worth the extra planning over one-at-a-time testing?
Whenever more than one or two factors plausibly interact, and getting the production setting wrong is costly. For a simple, well-understood single-variable relationship, one-at-a-time testing can be perfectly adequate.
What's the difference between full factorial and fractional factorial?
Full factorial tests every combination of every factor and level; fractional factorial tests a carefully chosen subset that still estimates the main effects, at far fewer runs, at the cost of some ability to distinguish certain interactions from each other.
How many runs does a DOE typically need?
It depends on the number of factors and the design chosen; a 2-level, 3-factor full factorial needs 8 runs, a 2-level, 5-factor full factorial needs 32, which is exactly why fractional designs and screening experiments exist for larger factor counts.
WRYGT uses DOE to optimize process windows during process development, not to chase a single good-looking sample lot. If a process needs real characterization before it scales,
talk to us about engineering services.