CHEN 4860 Unit Operations Lab Design of Experiments (DOE) With excerpts from “Strategy of Experiments” from Experimental Strategies, Inc.
DOE Lab Schedule
DOE Lab Schedule Details Lecture 2 Limitations of Factorial Design Centerpoint Design Screening Designs Response Surface Designs Formal Report
Limitations of Factorial Design Circumventing Shortcomings
Limitations of 2 k Factorials Optimum number of trials? “Signal-to-Noise” ratio Nonlinearity? 3 k factorial or center point factorial Inoperable regions? Tuck method Too many variables? Screening designs Fractional Factorial Plackett-Burman Need detailed understanding? Response Surface Plots
Number of Runs vs. Signal/Noise Ratio Confidence Interval or Signal FEavg + t*SeffFEavg - t*Seff FEavg + t*SeffFEavg - t*Seff
Number of Runs vs. Signal/Noise Ratio Avg + t*Seff = 2*t*Seff Seff = 2*Se/sqrt(N) = 2*2*t*Se/sqrt(N) Rearrange, N (total number of trials) is: N=[2*2*t/(/Se)]^2 Estimate t as approximately 2 N=[(7 or 8)/(/Se)]^2
Number of Runs vs. Signal/Noise Ratio (/Se) is the signal to noise ratio. Ratio (/Se) Number of Runs to to to to 256
Number of Runs vs. Signal/Noise Ratio
Factorial Design (2 k ) 2 is number of levels (low, high) What about non-linearity? A B C LO, HI, LO LO, LO, LO Pts (A, B, C) HI, LO, LO HI, HI, LO HI, LO, HI HI, HI, HI LO, HI, HI LO, HI, LO
Centerpoint Test for Nonlinearity Additional pts. located at midpoints of factor levels. (No longer 8 runs, Now 20) A B C LO, HI, LO LO, LO, LO Pts (A, B, C) HI, LO, LO HI, HI, LO HI, LO, HI HI, HI, HI LO, HI, HI LO, HI, LO
Centerpoint Test for Non-linearity Effect(nonlinearity) =Ynoncpavg-Ycavg What about significance? Calculate variance of non-centerpoint (cp) tests as normal (S^2) Calculate variances of cp (Sc^2) Degrees of Freedom (df) for base design (#noncp runs)*(reps/run-1) DF for cp (dfc) (#cp runs-1) Calculate weighted avg variance Se^2 = [(df*S^2)+(dfc*Sc^2)]/(dfc+df) Snonlin=Se*sqrt(1/Nnoncp+1/Ncp) dftot=dfc+df Lookup t from table using dftot Calculate DL = + t*Snonlin
Better Way to Test Non-Linearity Use response surface plots with Face Centered Cubes, Box-Behnken Designs, and others. Face-Centered Cube (15 runs)Box-Behnken Design (13 runs)
Inoperable Regions Don’t shrink design, pull corner inward X1 X2 X1 X2 BADGOOD
Diagnosing the Environment Too many variables, use screening designs to pick best candidates for factorial design Screening Designs Full Factorial Designs Response Surface Designs Many Independent Variables Fewer independent variables (<5) “Crude” Information Quality Linear Prediction Quality non-linear Prediction
Screening Designs Benefits: Only few more runs than factors needed Used for 6 or more factors Limitations: Can’t measure any interactions or non- linearity. Assume effects are independent of each other
Screening Designs # of runs needed # of FactorsFull FactorialScreening Design , , , ,388, ,217,72828
Screening Designs Fractional Factorial Interactions are totally confounded with each other in identifiable sets called “aliases”. Available in sizes that are powers of 2. Plackett-Burman Interactions are partially correlated with other effects in identifiable patterns Available in sizes that are multiples of 4.
Fractional Factorial (1/2-Factorial) Suppose we want to study 4 factors, but don’t want to run the 16 experiments (or 32 with replication). ABCABACBCABC Typical Full Factorial
Fractional Factorial What happens if we replace the unlikely ABC interaction with a new variable D? The other 2 factor interactions become confounded with one another to form “aliases” AB=CD, AC=BD, AD=BC The other 3 factor interactions become confounded with the main factor to also form “aliases” A=BCD, B=ACD, C=ABD
Fractional Factorial Ignoring the unlikely 3 factor interaction, we have… ABCDAB=CDAC=BDAD=BC
Fractional Factorial Calculations performed the same If the effects of interactions prove to be significant, perform a full factorial with the main effects to determine which interaction is most important.
Plackett-Burman Benefits: Can study more factors in less experiments Costs: Main factor in confounded with all 2 factor interactions. Suppose we want to study 7 factors, but only want to run 8 experiments (or 16 with replication).
Plackett-Burman A=BD= CG=EF B=AD= CE=FG C=AG= BE=DF D=AB= CF=EG E=AF= BC=DG F=AE= BG=CD G=AC= BF=DE
Plackett-Burman Calculations performed the same How do you handle confounding of main affects? Use General Rules: Heredity: Large main effects have interactions Sparsity: Interactions are of a lower magnitude than main effects Process Knowledge Use Reflection
Reflection of Plackett-Burman Reruns the same experiment with the opposite signs. A=BD= CG=EF B=AD= CE=FG C=AG= BE=DF D=AB= CF=EG E=AF=B C=DG F=AE=B G=CD G=AC= BF=DE
Reflection of Plackett-Burman Treats 2 factor responses as noise Average the effects from each run to determine the true main effect Normal E(A)calc=E(A)act-Noise Reflected E(A)calcr=E(A)actr+Noise Combined E(A)est=(E(A)calc+E(A)calcr)/2
Response Surface Plots Need detail for more than 1 response variable and related interactions Types 3 level factorial Face-Centered Cube Design Box-Behnken Design Many experiments required
Size of Response Surface Design Number of Factors 3-level Factorial Face- Centered Cube* Box- Behnken* *extra space left for multiple center points due to blocking
Summary Diagnose your problem Use one of the many different methods outlined to circumvent it Many more options and designs listed on the web
Formal Memo Follow outline presented for formal memo presented on Dr. Placek’s website. Executive Summary Discussion and Results Appendix with Data, Calcs, References, etc. **GOAL IS PLANNING**
Formal Memo Report Questions What are your objectives? How did you minimize random and bias error? What variables did you control and why? What variables did you measure and why? What were the results of your experiment? Which factors were most important and why? What is your theory (based on chem-eng knowledge) on why the experiment turned out the way it did? Was there any codependence? What will be your next experiment? What would you do differently the next time?