Design of Experiments (DOE) is a structured statistical method that varies several process factors together, in a planned pattern, so you can measure each factor's effect and how factors interact, all from a compact set of trial runs. Instead of changing one thing at a time and hoping the result holds, DOE lets you learn the most about your process from the fewest experiments. In manufacturing it is the workhorse of the Improve phase in DMAIC, where teams tune settings to cut defects, lift yield, and stabilize output.
The intuitive approach is OFAT: hold everything fixed, change one factor, record the result, then move to the next factor. It feels rigorous but has two fatal weaknesses. First, it is inefficient, because each factor needs its own separate block of runs. Second, and more damaging, OFAT is blind to interactions, the situations where the effect of one factor depends on the level of another. DOE varies factors simultaneously in a balanced pattern, so a single experiment reveals both individual effects and the combined effects that OFAT can never see.
Three terms define every experiment. A factor is an input you can control, such as oven temperature, feed rate, or dwell time. A level is a specific setting you test for that factor, for example 180C versus 200C. The response is the measured output you care about, such as tensile strength, cycle time, or scrap rate. A design that studies two factors, each at two levels, is called a 2x2 (or two-level, two-factor) design and produces four unique factor combinations to run.
A full factorial design tests every possible combination of factor levels. For two-level designs the run count is 2 raised to the number of factors: 2 factors need 4 runs, 3 factors need 8, but 7 factors would balloon to 128 runs. When that becomes impractical, a fractional factorial tests a carefully chosen subset (a half, quarter, or eighth) that still estimates the main effects and the most important interactions. The tradeoff is confounding: some higher-order interactions get aliased together, so you trade a little resolution for a large saving in time and material.
If you want to go deeper into robust design, Taguchi methods build on the same fractional factorial logic, using standardized orthogonal arrays to keep experiments small while making the process less sensitive to noise.
A main effect is the average change in the response when a factor moves from its low level to its high level, ignoring the other factors. An interaction effect measures how much a factor's influence shifts depending on another factor's setting. Interactions are where the real money often hides: a coating that only performs well at high temperature and high pressure would never surface in OFAT testing. Knowing which effects dominate lets you prioritize your improvement work, much as Pareto analysis ranks the vital few causes over the trivial many.
Suppose you want to raise the tensile strength (in megapascals) of a molded part. You choose two factors: temperature (A) at 180C (low) and 200C (high), and pressure (B) at 40 bar (low) and 60 bar (high). The four runs of the full factorial give these responses:
Main effect of A (temperature): average the two high-A runs and subtract the average of the two low-A runs. High-A average = (60 + 80) / 2 = 70. Low-A average = (50 + 54) / 2 = 52. Main effect of A = 70 minus 52 = 18 MPa. Raising temperature adds about 18 MPa on average.
Main effect of B (pressure): high-B average = (54 + 80) / 2 = 67. Low-B average = (50 + 60) / 2 = 55. Main effect of B = 67 minus 55 = 12 MPa.
Interaction effect (A x B): compute the effect of A when B is high, then when B is low, and take half the difference. Effect of A at high B = 80 minus 54 = 26. Effect of A at low B = 60 minus 50 = 10. Interaction = (26 minus 10) / 2 = 8 MPa. That positive interaction means temperature and pressure reinforce each other: pushing both to high yields far more than the two main effects would predict alone. An OFAT study would have missed it entirely.
Two design principles keep your conclusions honest. Replication means running each combination more than once, so you can separate a real effect from random measurement scatter and estimate experimental error. Randomization means running the trials in random order rather than a tidy sequence, so slow drifts (a warming shop floor, a tiring operator, tool wear tracked by the bathtub curve) do not masquerade as a factor effect. Together they let you trust that a measured difference came from the factor and not from confounded background variation. This is the same discipline that underpins a sound control plan once the winning settings are locked in.
Fabrico does not run the DOE analysis for you, and it is not statistical software. What it does is supply the accurate, timestamped production data that any experiment depends on. Fabrico delivers real-time OEE and production monitoring, including computer-vision monitoring that captures cycle counts, stops, and output even on machines without a PLC. During a designed experiment that means your response values (yield, cycle time, defect counts) are logged automatically instead of hand-recorded, so the numbers feeding your main-effect and interaction calculations are clean. Combined with a robust measurement system checked through Gauge R&R, this removes a major source of error before it ever reaches your model. After the experiment, Fabrico's CMMS helps you sustain the improved settings and monitor the process with statistical process control and process capability studies. DOE also pairs naturally with an FMEA, which points to the highest-risk factors worth putting into your design.
Technically many, but practically you should screen first. With more than four or five factors, teams use a fractional factorial screening design to identify the two or three factors that matter most, then run a focused full factorial or response-surface design on those. This keeps run counts manageable while still surfacing the dominant main effects and key interactions.
Simple two-level, two-factor experiments can be calculated by hand, as the worked example above shows. Larger designs are usually built and analyzed in a statistics package because the arithmetic for confounding, blocking, and significance testing grows complex. Fabrico is not that software; it provides the reliable production data those tools analyze, so your inputs are trustworthy from the start.
DOE is an active, one-time investigation: you deliberately change settings to discover cause-and-effect relationships and find the best configuration. SPC is passive and ongoing: it monitors a running process to detect when it drifts out of control. You use DOE to reach a good state, then use SPC and capability analysis to hold it there.
Ready to run experiments on data you can trust? Fabrico gives you the real-time, automatically logged production numbers that make every DOE, SPC chart, and capability study honest. Book a Fabrico demo to see the data foundation behind better decisions.
Login
Blog