Why DOE Matters
Most process optimization on the shop floor follows a one-factor-at-a-time (OFAT) approach: change one variable, hold everything else constant, see what happens. It feels logical but it is dangerously incomplete. OFAT cannot detect interactions — situations where factor A and factor B together produce an effect that neither produces alone. In manufacturing, interactions are everywhere: temperature and pressure, speed and feed rate, cure time and humidity. DOE tests multiple factors simultaneously in a structured way, revealing both main effects and interactions in far fewer runs than OFAT.
| Approach | Factors Tested | Finds Interactions? | Runs Required (3 factors, 2 levels) |
|---|---|---|---|
| OFAT | One at a time | ❌ No | 6+ runs (and still misses interactions) |
| Full Factorial | All combinations | ✅ All of them | 8 runs (2³) |
| Fractional Factorial | Strategic subset | ✅ Most important | 4 runs (half-fraction) |
Key DOE Terminology
| Term | Definition | Welding Example |
|---|---|---|
| Factor | A process variable you control | Weld speed, voltage, wire feed rate |
| Level | The specific settings tested for a factor | Speed: 20 ipm (low) vs. 30 ipm (high) |
| Response | The output you measure | Weld tensile strength (MPa) |
| Main Effect | The average impact of changing one factor | Increasing voltage raises strength by 15 MPa |
| Interaction | Combined effect of two+ factors that differs from their individual effects | High voltage + low speed boosts strength far more than either alone |
| Replicate | Repeating the entire experiment to estimate error | Run all 8 combinations twice = 16 total runs |
| Randomization | Running trials in random order to avoid lurking variables | Do not run all "high speed" trials in the morning |
Main Effects vs. Interactions
A main effect tells you the average impact of a single factor across all runs. An interaction tells you that the effect of one factor depends on the level of another. Interactions are the reason OFAT fails — and the primary reason DOE exists.
Why Interactions Matter More Than You Think
In a review of over 100 published industrial experiments, more than 75% had at least one significant interaction. If you optimize factors one at a time, you will likely settle on settings nowhere near the true optimum — because you never saw how the factors work together. DOE is the only systematic way to uncover these joint effects.
Reading an interaction plot: If the lines are parallel, there is no interaction — the effect of Factor A is the same regardless of Factor B. If the lines are not parallel (they converge, diverge, or cross), there is an interaction. Crossing lines indicate a strong interaction where the best level of one factor flips depending on the other.
2k Factorial Design — Welding Example
A fabrication shop needs to optimize MIG weld tensile strength. The team selects three factors at two levels each, giving a 2³ = 8 run experiment:
| Run | Speed (ipm) | Voltage (V) | Wire Feed (in/min) | Strength (MPa) |
|---|---|---|---|---|
| 1 | Low (20) | Low (22) | Low (150) | 285 |
| 2 | High (30) | Low (22) | Low (150) | 270 |
| 3 | Low (20) | High (28) | Low (150) | 310 |
| 4 | High (30) | High (28) | Low (150) | 325 |
| 5 | Low (20) | Low (22) | High (250) | 295 |
| 6 | High (30) | Low (22) | High (250) | 290 |
| 7 | Low (20) | High (28) | High (250) | 340 |
| 8 | High (30) | High (28) | High (250) | 360 |
Analysis: Voltage has the largest main effect (+32.5 MPa average). The Speed × Voltage interaction is significant — high voltage combined with high speed gives a disproportionate boost that neither factor produces alone. An OFAT study testing speed in isolation would have concluded speed hurts strength and set it low, missing the powerful synergy at high voltage.
Full Factorial vs. Fractional Factorial
Full factorials test every combination but grow exponentially. With 6 factors at 2 levels, you need 64 runs. Fractional factorials strategically skip combinations to cut runs while still estimating the most important effects:
| Factors | Full Factorial Runs | Half-Fraction Runs | What You Sacrifice |
|---|---|---|---|
| 3 | 8 | 4 | 3-factor interaction (rarely significant) |
| 4 | 16 | 8 | Some 2-factor interactions aliased |
| 5 | 32 | 16 | Higher-order interactions aliased |
| 6 | 64 | 16 (quarter-fraction) | Some 2-factor interactions confounded |
The Sparsity Principle
In practice, most process variation is explained by main effects and two-factor interactions. Three-factor and higher interactions are rarely significant. This is why fractional factorials work — you sacrifice information about effects that almost never matter. Use Resolution IV or higher designs to keep main effects clean from two-factor interaction aliasing.
Response Surface Methodology (RSM)
Once screening experiments identify the vital few factors, RSM helps you find the optimal settings. RSM uses center points and additional levels to fit a curved (quadratic) model, revealing the peak or valley in the response surface.
Central Composite Design (CCD) adds axial (star) points outside the factorial range, creating 5 levels per factor — the workhorse RSM design. Box-Behnken Design avoids extreme corner combinations, which is useful when those settings are dangerous or impossible to run. Both generate a contour plot showing how the response changes across two factors simultaneously — a visual map to your optimum.
Running a DOE on the Shop Floor
Analysis Tools & Interpretation
Minitab and JMP are the industry standards for DOE analysis. Both generate three critical outputs every IE must learn to read:
| Output | What It Shows | What to Look For |
|---|---|---|
| Main Effects Plot | Average response at each factor level | Steeper slope = bigger effect. Flat line = factor does not matter. |
| Interaction Plot | How one factor's effect changes at different levels of another | Non-parallel lines = significant interaction. Crossing lines = strong interaction. |
| Pareto of Effects | Bar chart ranking all effects by magnitude | Bars crossing the significance line are statistically significant (p < 0.05). Focus resources there. |
The Confirmation Run — Never Skip It
Run 3–5 trials at the predicted optimal settings and verify the response matches the model prediction within confidence bounds. If it does not match, you likely have an uncontrolled noise factor the model did not capture. Investigate before rolling changes to production.
When DOE Is Essential vs. Overkill
✅ Use DOE When
- Multiple factors may interact (welding, injection molding, heat treat)
- OFAT tweaking has stalled — you cannot improve further by gut feel
- New process launch needs optimized parameters fast
- Cpk is below target and you need to find better settings
- Customer or validation requires statistical evidence of optimization
❌ Skip DOE When
- Root cause is obvious (broken part, wrong material loaded)
- Only one factor can realistically be changed (no design space)
- Process is unstable — get it in control first
- Measurement system is unreliable (fix gauge R&R before experimenting)
- You lack authority or budget to actually change the process settings
Common DOE Mistakes & How to Avoid Them
🎯 Key Takeaway
Design of Experiments replaces trial-and-error guessing with structured, data-driven optimization. It reveals interactions that OFAT will never find, identifies the vital few factors from the trivial many, and reaches the optimum in far fewer runs. Start with a clear objective and a measurable response, use a screening factorial to narrow your factors, then RSM to dial in the sweet spot. Randomize every time, always run confirmation trials, and update your standard work and SPC charts to lock in the gains. DOE is the most powerful optimization tool in the IE toolkit — learn it, use it, and stop guessing.
Interactive Demo
Run a 2-factor factorial experiment. Set factor levels, observe the response at each combination, and see which factors and interactions have the biggest effects.
Stop reading, start modeling
Model your process flow, run simulations, optimize staffing with TOC math, and test your knowledge with 107 interactive checks — all in one platform.