Practical Application of Design of Experiments (DOE): Three Case Studies from Workshop Confusion to Process Breakthrough

1. Introduction: Why DOE is a Core Tool in 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? The traditional one-factor-at-a-time (OFAT) approach is not only inefficient but also fails to identify interactions between factors—these interactions are often the key to process breakthroughs.

Design of Experiments (DOE) is a statistical methodology for systematically planning and analyzing experimental data. It was pioneered by British statistician R.A. Fisher in the 1920s and has since been further developed by Box, Taguchi, and others. DOE is one of the most powerful tools in the Six Sigma Define-Measure-Analyze-Improve-Control (DMAIC) improvement phase.

However, many quality professionals, after learning the theory of DOE, are most puzzled by 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 line issues, they often don't know where to start—too many factors can lead to an explosion in the number of trials, while too few factors may miss critical ones. Data analysis can also be overwhelming, with a multitude of p-values and effect plots making it difficult to make decisions.

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

2. Case Study 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 nonconformity rate as high as 8.3%. The customer demanded a corrective action plan within 48 hours, or the supplier would face the risk of a production stoppage, with monthly penalties reaching up to 3% of sales.

The initial team adjusted two parameters, mold temperature and holding pressure, based on experience. After four trials, the warpage rate decreased from 8.3% to 6.5%, but it could not be reduced further. More troubling was that when another injection molding machine was used, 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 adjustments 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 "robbing Peter to pay Paul" phenomenon is a typical signal of interactions between multiple factors.

Factor Screening and Design

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

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

Execution and Data Analysis

The 16 trials were randomly conducted 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 key finding: when the mold temperature was above 50°C, 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°C.

Optimization Results

Using the response optimizer, the team found the optimal parameter combination: mold temperature 55°C, melt temperature 235°C, 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, a standard deviation of 0.06mm, and a Cpk improvement from 0.67 to 1.53, reducing the nonconformity rate to 0.2%.

Insights

Insight: 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 root cause of the problem. More importantly, the traditional OFAT method would have led them to believe they had found the "optimal solution" (a 6.5% nonconformity rate might have been deemed acceptable), when in fact they were far from the true potential of the process.

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

Background and Pain Points

An EMS contract manufacturer in Shenzhen encountered an issue with excessive void rates in BGA (ball grid array) solder joints during the mass production of a new smartphone motherboard. The IPC standard requires a void rate of ≤25%, but the product's average void rate was 31.2%, with some joints exceeding 45%. Even more challenging was the fact that the high void rate was not consistent across all 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, adding to the difficulty of root cause analysis.

The engineering team suspected that the temperature profile of the reflow soldering process 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 approach is to refer to the recommended temperature profile from the solder paste supplier, 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 (CCD)

Through preliminary screening trials (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 a total of 50 trials (including 6 center points).

Key Findings

After fitting a second-order model, the analysis of variance (ANOVA) showed:

1. Reflow Peak Temperature (C) was the most significant single factor affecting the void rate (42% contribution). Higher peak temperatures reduced void rates, but the improvement slowed after 250°C.
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 the void rate decreased.
3. The 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°C, soak zone 182°C, reflow peak 248°C, cooling slope 2.8°C/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.

Insight: CCD saved about 60% of the trial runs compared to a full factorial design, while still being able to fit curvature effects and identify the true extremum. This two-stage strategy of "screening first, then optimizing" is a standard paradigm in the practical application of DOE.

4. Case Study Three: Optimization of Catalyst Formulation at a Chemical Company

Background and Pain Points

This is a fine chemical company that produces polyolefin catalysts. The catalytic activity of their main product (measured in kg product/g catalyst) 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—meaning the current parameter window might have already deviated from the true "peak."

Approach: Response Surface Methodology (RSM)

The project team adopted 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 quadratic 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 plot and response surface plot clearly showed that the optimal region was at X1≈+0.3 (calcination temperature about 15°C 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, respectively, averaging 917, reaching the industry benchmark level. Based on a selling price of 120,000 RMB per ton of catalyst, the improvement in activity translated to an annualized benefit of about 6.3 million RMB. More importantly, this optimization significantly improved batch-to-batch stability—before optimization, the activity fluctuated by more than 150 (850～700), but after optimization, it stabilized between 908 and 926, reducing the variation by 83%.

Insights

Insight: 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 moving further away from the optimal solution. This case also highlights that the value of DOE lies not only in finding the optimum but also in improving process consistency—many quality professionals focus on improving the mean, overlooking the hidden benefits of reducing variance. 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, efforts need to be made in the following areas:

1. Preparatory Work is More Important than the Experiment Itself

In Case One, the team spent two days on cause-and-effect matrix analysis to narrow down the factors from twelve to five. 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 that 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 in 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 the Ability to Detect Interactions

In Case One, the team chose a 2^(5-1) design with Resolution V rather than III to detect second-order interactions. While a Resolution III design requires fewer trials, it confounds main effects with interactions, potentially 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 the statistical model 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 receives basic DOE training, and for problems involving three or more factors, DOE must be conducted before machine adjustments. This system has reduced the average problem resolution 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 lab. Its core value is not in "conducting experiments" but in using the least resources and the most scientific methods to uncover the true relationship between factors and responses. When a quality engineer masters DOE, they are no longer a trial-and-error adjuster but a systematic experimenter who can speak with data.

These three cases cover injection molding, electronics assembly, and chemical processes in manufacturing, but the application of DOE extends far beyond these areas—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 best solution

Knowledge Number: 6.4.1

Version: v20260719

Author: Quality Excellence Think Tank The Quality Excellence Think Tank is dedicated to providing systematic professional knowledge, methodologies, and practical tools for quality management practitioners, supporting continuous improvement in quality capabilities within enterprises.