Design of Experiments (DOE) in Practice: Three Real-World Cases from the Shop Floor to Process Breakthroughs

1. Introduction: Why DOE is the Core Tool of Six Sigma

In quality improvement efforts within manufacturing and service industries, we often face a common challenge: with so many factors (X) influencing the output (Y), which ones are truly critical? Traditional one-factor-at-a-time (OFAT) methods are not only inefficient but also fail to identify interactions between factors—interactions that are often the key to process breakthroughs.

Design of Experiments (DOE) is a statistical methodology that systematically arranges experimental designs and scientifically analyzes data. It was pioneered by British statistician R.A. Fisher in the 1920s and further developed by Box, Taguchi, and others. DOE has become one of the most powerful tools in the Six Sigma DMAIC improvement phase.

However, many quality professionals, after learning the theory of DOE, are often confused about how to apply it in actual production. Concepts like full factorial design, fractional factorial design, and response surface methodology (RSM) are well understood in the classroom, but when faced with real production issues, they often struggle to start—too many factors can lead to an explosion in the number of trials, while too few factors might miss critical ones. Data analysis can also be challenging, with a plethora of p-values and effect plots to interpret.

Today, we will use three real-world cases from different industries to illustrate the entire process of applying DOE, from problem definition to implementation. These cases represent three typical scenarios of DOE application: factor screening and interaction analysis (injection molding), surface optimization and process window identification (reflow soldering), and formula optimization and multi-factor optimization (catalyst development). We hope these cases will show that DOE is not just the domain of statisticians but a practical tool that every engineer and quality professional can master.

2. Case One: Optimization of Injection Molding Process at an Automotive Parts Manufacturer

Background and Pain Points

In 2023, an automotive interior component supplier in East China received a customer complaint: the dashboard trim strips they produced exhibited batch-wise warpage after assembly, with a defect rate as high as 8.3%. The customer demanded a corrective action plan within 48 hours, or the supplier would face a production halt and a monthly penalty of up to 3% of sales.

The initial team adjusted two parameters, mold temperature and holding pressure, and conducted four trials, reducing the warpage rate from 8.3% to 6.5%. However, they could not reduce it further. More troubling, when they switched to another injection molding machine, the same parameter combination resulted in a new deformation pattern, leading the team to suspect machine differences. The on-site engineers repeatedly adjusted the machine settings, but the results became increasingly chaotic—increasing the mold temperature reduced warpage but caused sink marks, while decreasing the holding pressure eliminated sink marks but increased warpage. This "pushing down one end and the other pops up" phenomenon is a typical signal of interactions between multiple factors.

Factor Screening and Plan Design

The project team decided to use a fractional factorial design (2^(5-1) Resolution V) for screening. Through brainstorming and a cause-and-effect matrix analysis, they identified five key factors from 12 potential factors:

The response variable Y was the warpage deformation at a specified measurement point (mm), with a target value of ≤0.5mm.

Execution and Data Analysis

The 16 trials were randomly executed on two injection molding machines, with each condition repeated three times. The Minitab analysis results were as follows:

• Significant Main Effects: A (mold temperature), B (melt temperature), and C (holding pressure) all had p-values <0.01.
• Significant Second-Order Interactions: A×C (mold temperature × holding pressure) p=0.003, B×E (melt temperature × cooling time) p=0.015.
• Model R²=94.7%, adjusted R²=92.3%.

The interaction plots revealed a critical finding: when the mold temperature was above 50℃, the effect of holding pressure on warpage decreased; in the low-temperature region, holding pressure was decisive. This means that the mold temperature should not be allowed to drop below 50℃.

Optimization Results

Using the response optimizer, the team found the optimal parameter combination: mold temperature 55℃, melt temperature 235℃, holding pressure 85 bar, holding time 10s, and cooling time 20s. A verification batch of 500 pieces was produced, with an average warpage deformation of 0.32mm and a standard deviation of 0.06mm. The CPK improved from 0.67 to 1.53, and the defect rate dropped to 0.2%.

Insights

Insights: Without DOE, the team would have continued to trial and error on the mold temperature factor, never realizing that its interaction with holding pressure was the true root cause. More importantly, the traditional OFAT method would have led them to believe they had found the "optimal solution" (a 6.5% defect rate might be acceptable), but in reality, they were far from the process's true potential.

3. Case Two: Yield Improvement in SMT Reflow Soldering at an Electronics Manufacturer

Background and Pain Points

A Shenzhen-based EMS contract manufacturer encountered an issue with excessive void rates in BGA (ball grid array) solder joints while mass-producing a new smartphone motherboard. The IPC standard requires a void rate of ≤25%, but the average void rate for this product was 31.2%, with some points exceeding 45%. More challenging was the fact that the high void rate was not consistent across all solder joints but showed clear cyclical variations—products from the day shift were better than those from the night shift, and products from Monday were better than those from Friday, further complicating the root cause analysis.

The engineering team suspected that the reflow soldering temperature profile was the issue, but a reflow soldering oven has 8-10 temperature zones, each of which can be independently set, leading to nearly infinite combinations. The standard practice is to refer to the solder paste supplier's recommended profile, but this profile is designed for standard boards and is not suitable for this high-density 12-layer board.

Factor Screening and Central Composite Design

Through preliminary screening experiments (Plackett-Burman design, 12 trials), the team identified four critical temperature zones and added one process factor—nitrogen flow rate:

A central composite design (CCD) was used, requiring 50 trials (including 6 center points).

Key Findings

After fitting a second-order model, the ANOVA results showed:

1. Reflow Peak Temperature (C) was the most significant single factor affecting the void rate (contribution rate 42%). Higher peak temperatures reduced void rates, but the improvement slowed after 250℃.
2. The interaction between Nitrogen Flow Rate (E) and Soak Zone Temperature (B) contributed 18%—at high nitrogen flow rates, the sensitivity of the soak zone temperature to void rates decreased.
3. The model's curvature effect was significant, indicating the presence of an optimal region that a simple first-order model could not capture.

The final optimized soldering conditions were: preheat zone 168℃, soak zone 182℃, reflow peak 248℃, cooling slope 2.8℃/s, and nitrogen flow rate 25 L/min. The verification batch showed an average void rate of 9.7% and a maximum of 18.3%, fully meeting the IPC standard.

Insights: CCD saved about 60% of the trial runs compared to a full factorial design while still fitting the curvature effect and identifying the true extremum point. This "screening first, then optimizing" two-stage strategy is a standard paradigm in the practical application of DOE.

4. Case Three: Catalyst Formula Optimization at a Chemical Company

Background and Pain Points

This is a specialty chemical company that produces polyolefin catalysts. The catalytic activity (units: kg product/g catalyst) of their main product had been declining for the past six months, from a baseline of 850 to 620, far below the industry benchmark of 900+. Laboratory trials indicated that there was room for optimization in three key parameters of the catalyst preparation process:

• Carrier Calcination Temperature (X1)
• Active Component Impregnation Concentration (X2)
• Reduction Treatment Time (X3)

However, the issue was that these three factors might have complex interactions, and the team suspected that the true optimal solution might lie outside the tested parameter range—i.e., the current parameter window might have already deviated from the true "peak."

Plan: Response Surface Methodology (RSM)

The project team used a Box-Behnken design, with three levels for each of the three factors, totaling 15 trials (including 3 center points). Compared to CCD, Box-Behnken requires fewer trials in a three-factor scenario and does not require running at extreme conditions—this is crucial for the safety and cost of chemical production.

Analysis and Verification

The fitted second-order regression model was:

Y = 893 + 47.2X1 + 38.5X2 + 21.8X3 - 81.3X1² - 64.7X2² - 29.5X3² + 18.6X1X2 + 9.2X1X3 - 12.4X2X3

R²=97.8%, lack-of-fit test p=0.21 (not significant, indicating a good model fit).

The contour and response surface plots clearly showed that the optimal region was located at X1≈+0.3 (calcination temperature about 15℃ higher than the current setting), X2≈+0.2 (impregnation concentration about 0.8% higher), and X3 around 0 (reduction time unchanged). Interestingly, the team's original operating point (X1=-0.5, X2=-0.3, X3=0) was at the "base" of the response surface—adjusting towards lower temperatures based on experience only led them further away from the optimal solution.

The verification trials were repeated in three batches, with catalytic activities of 917, 926, and 908, averaging 917, reaching the industry benchmark level. At a selling price of 120,000 yuan per ton of catalyst, the increased activity translates to an annualized benefit of approximately 6.3 million yuan. More importantly, this optimization significantly improved batch consistency—the activity range before optimization was over 150 (850～700), and after optimization, it stabilized between 908 and 926, reducing the variation by 83%.

Insights

Insights: Intuition and experience are not always reliable, especially in multi-dimensional parameter spaces. Response surface methodology (RSM) not only helped the team find the right direction but also prevented them from continuing down the wrong path. This case also highlights that the value of DOE lies not only in finding the optimal solution but also in improving process consistency—many quality professionals focus on improving the mean, but the reduction in variance can bring significant hidden benefits. In fact, reducing process variation often has a greater impact on long-term quality costs than improving the mean.

5. Key Points for Successful DOE Implementation

Summarizing the three cases, to truly leverage DOE in practical applications, the following aspects need to be addressed:

1. Preparatory Work is More Important than the Experiment Itself

In Case One, the team spent two days on a cause-and-effect matrix analysis to narrow down 12 factors to 5. Selecting the wrong factors can render even the best experimental design ineffective. It is recommended to use a progressive screening process: fishbone diagram → cause-and-effect matrix → FMEA, to ensure no critical factors are overlooked.

2. Randomization and Blocking Cannot Be Skipped

In Case Two, if the 50 trials were conducted in batches based on convenience, environmental drifts such as temperature and humidity would contaminate the data. Randomizing the trial sequence and introducing blocks when necessary are prerequisites for valid statistical inference.

3. Resolution Determines Whether Interactions Can Be Identified

In Case One, the team chose a 2^(5-1) design with Resolution V rather than III to identify second-order interactions. While Resolution III designs have fewer trials, they confound main effects with interactions, leaving potential issues unresolved and leading to higher rework costs in later stages.

4. Verification Trials Are Not Optional

All three cases conducted independent verification batches after optimization. The "optimal solution" provided by statistical models is an estimate based on limited trial data and must be confirmed with new data. In Case Three, the verification batch results were very close to the predicted values (predicted 908, actual 917), demonstrating the high reliability of the model.

5. Integrate DOE into the Company's Problem-Solving Process

The most successful approach is not a single DOE project but integrating DOE into the standard steps of engineering changes. For example, an automotive parts manufacturer's "3+1" process: problem description → DOE screening → DOE optimization → verification and solidification. Every engineer has received basic DOE training, and any problem involving three or more factors must be addressed with DOE before machine adjustments—this system reduced the average problem-solving cycle from 23 days to 5 days.

6. Conclusion

DOE is not a statistical toy confined to ivory towers but a practical tool for solving complex problems on the shop floor and in the laboratory. Its core value is not in "conducting experiments" but in using the least resources and the most scientific methods to find the true relationship between factors and responses. When a quality engineer masters DOE, they are no longer a trial-and-error experimenter but a systematic experimenter who can speak with data.

The three cases cover injection molding, electronics assembly, and chemical processes in manufacturing, but the application of DOE extends far beyond these fields—any scenario in healthcare, pharmaceuticals, food, and financial services where multiple variables influence outcomes can benefit from DOE methodology. Mastering DOE is like holding the key to unlocking complex problems.

Use the fewest experiments to find the optimal solution

Knowledge code: 6.4.1

Version: v20260719

Author: Quality Think Tank Quality Think Tank is dedicated to providing systematic knowledge, methodologies, and practical tools for quality management professionals, helping companies continuously improve their quality capabilities.