Quote:
Originally Posted by speeder
Well, he can write a sentence in the English language, so there is that.
I know nothing about engineering or even what a DOE is but I know English when I see it, or don't.
|
Anyone that is in R&D or engineering that doesn't do designed experiments will waste a ton of time. These techniques are normally used with high-end development and refinement. A designed experiment allows you to look at several different variables without having to do a ton of trials. Here's a simplified decision tree to guide you in selecting a designed experiment:
1. Determine your research objective:
• Is it screening and identifying key factors?
• Are you looking to estimate the main effects and interactions?
• Do you want to build a response surface model and optimize the process?
2. Consider the number of factors involved:
• One factor: One-Factor-at-a-Time (OFAT): Simple but inefficient for exploring interactions.
• Two factors:
o Full Factorial: Efficient for exploring all combinations and interactions but can be costly for many factors.
o 2-Factor Central Composite Design (CCD): More advanced, allows for model building and optimization with curvature exploration.
• Three or more factors:
o 3-Factor Central Composite Design (CCD): More advanced, allows for model building and optimization with curvature exploration.
o Box-Behnken Design: Efficient for exploring quadratic terms without requiring as many runs as a full factorial.
o Plackett-Burman Design: Useful for screening many factors with limited resources, but only provides information about main effects.
o Fractional Factorial Design: Efficient for screening and identifying key factors, requiring fewer runs than a full factorial.
o Derringer Design: Useful for optimizing multiple responses simultaneously when interactions are important.
3. Analyze your budget and resource constraints:
• Limited resources: Consider Plackett-Burman, Fractional Factorial, or even OFAT if interactions are not a major concern. Choose smaller designs with fewer runs.
• Ample resources: Full factorial, CCD, or Box-Behnken designs can be beneficial for detailed analysis and model building.
4. Assess the expected relationship between factors and response:
• Linear relationship: Factorial, Box-Behnken, or Fractional Factorial designs might suffice.
• Non-linear relationship: Central Composite Design (Derringer Design) might be better to capture curvature and complex interactions.
• Unknown relationship: Start with a Fractional Factorial design (Plackett-Burman design) to identify key factors, then follow up with a more specific design based on the findings.
5. Consider experimental error and accuracy needs:
• High accuracy: Include replicates in your design, especially at the center point. Choose designs with inherent replication or error estimation capability.
• Initial exploration and rough estimates: Fewer replicates might be acceptable, depending on the tolerance for error.