partition experimental designs for sequential process steps: application to product development...
Embed Size (px)
TRANSCRIPT

Partition Experimental Designs for Sequential Process Steps: Application to Product Development
Leonard Perry, Ph.D., MBB, CSSBB, CQE
Associate Professor & ISyE Program Chair
Industrial & Systems Engineering (ISyE)
University of San Diego
1

Example: Lens Finishing Processes
A company desires to improve their lens finishing process. Experimental runs must be limited due to cost concerns.
What type of design do you recommend?
2
ManufacturingProcess #1
Controllable factors
Uncontrollable factors
Inputs
Outputs, y
x1 x2 xk
z1 z2 zr
. . .
. . .
ManufacturingProcess #2
Controllable factors
Uncontrollable factors
Inputs
Outputs, y
x1 x2 xk
z1 z2 zr
. . .
. . .
Process One:Four Factors
Process Two:Six Factors

Objective of Partition Designs
To create a experimental design capable of handling a serial process consisting of multiple sequential processes that possess several factors and multiple responses.
Advantages: Output from first process may be difficult to
measure. Potential interaction between sequential
processes Reduction of experimental runs
3

Partition Design
4
ManufacturingProcess #1
Controllable factors
Uncontrollable factors
Inputs
Outputs, y
x1 x2 xk
z1 z2 zr
. . .
. . .
ManufacturingProcess #2
Controllable factors
Uncontrollable factors
Inputs
Outputs, y
x1 x2 xk
z1 z2 zr
. . .
. . .
x1 x21 1-1 11 -1-1 -11 1
x3 x41 1-1 11 11 -1-1 -1
R1 R234.43 12.419.94 2.7514.695 32.35
-31.34 -18.513.37 -8.625
Design Matrix #1 Design Matrix #2 Responses
+ =

Partition Design: Assumptions
Process/Product Knowledge required Screening Experiment required Resources limited, minimize runs Sparsity-of-Effect Principle
5

Partition Design: Methodology
6
1. Perform Screening Experiment for Each Individual Process
2. Construct Partition Design3. Perform Partition Design
Experiment4. Perform Partition Design
Analysisa) Select Significant Effects for
Each Responseb) Build Empirical Model for
Each Responsec) Calculate Partition Interceptd) Select Significant Effects for
Intercept5. Build Final Empirical Model
Screening Experiments for each Process
Process 1
Process 2
Process n
Partition Design Analysis
Select Significant
Effects
Perform Partition Design
Experiment
Construction of
Partition Design
Build Empirical
Model
Calculate Partition Intercept
Select Significant
Effects
Build Final Empirical Model

Review: Experimental Objectives
Product/Process Characterization Determine which factors are most influential on the observed response. “Screening” Experiments Designs: 2k-p Fractional Factorial, Plackett-Burman Designs
Product/Process Improvement Find the setting for factors that create a desired output or response Determine model equation to relate factors and observed response Designs: 2k Factorial, 2k Factorial with Center Points
Product/Process Optimization Determine an operating or design region in which the important factors
lead to the best possible response. (Response Surface) Designs: Central Composite Designs, Box-Behnken Designs, D-optimal
Product/Process Robustness Explore settings that minimize the effects of uncontrollable factors Designs: Taguchi Experiments
7

Example: First-order Partition Design Two factors significant in each process
Total of k = 4 factors Potential Interaction between processes
Partition Design N = 5 runs (N = k - 1) (Saturated Design)
8

Step 1: Perform Screening Experiment Process 1: Significant Factors:
Factor A Factor B
Process 2: Significant Factors:
Factor C Factor D
9

Step 2: Construct Partition Design Partition Design: Design Criteria
First-order models Orthogonal D-optimal Minimize Alias Confounding
Second-order models D-efficiency G-efficiency Minimize Alias Confounding
10

Step 2: Construct Partition Design First-order Design (Res III or Saturated)
Orthogonal D-optimality Minimize Alias Confounding
11
Step 1 Step 2 Step 3 Step 4
x1 x2 x3 x4 1 1 -1 1 1 -1
-1 -1
x1 x2 x3 x4 1 1 1 -1 1 -1 1 -1 -1
-1 -1 1
x1 x2 x3 x4 1 1 1 1 -1 1 -1 1 1 -1 -1 -1
-1 -1 1 -1
x1 x2 x3 X4 1 1 1 1 -1 1 -1 1 1 -1 -1 -1 -1 -1 1 -1 -1 -1 1 1

Step 2: Construct Partition Design
12
Term AliasesModel A-A BD CD ABCModel B-B AB BC BD ABC ABD BCDError C-C AB AD BC BD ABC ABD BCDError D-D AB AC BCD
X'X= 5 1 1 1 1(5x5) 1 5 1 1 1
Det(X'X)= 1 1 5 -3 11024 1 1 -3 5 1
1 1 1 1 5 Est. AB AC AD BC BD CDInt 0.0 0.0 0.0 -1.0 0.0 0.0
inv(X'X)= 0.25 0.00 -0.13 -0.13 0.00 A 0.0 0.0 0.0 0.0 -1.0 1.0(5x5) 0.00 0.25 -0.13 -0.13 0.00 B 1.0 0.0 0.0 1.0 1.0 0.0
-0.13 -0.13 0.50 0.38 -0.13 C 1.0 0.0 1.0 1.0 1.0 0.0Det(X'X)= -0.13 -0.13 0.38 0.50 -0.13 D -1.0 1.0 0.0 0.0 0.0 0.00.000977 0.00 0.00 -0.13 -0.13 0.25
Alias Matrix=inv(X'X)X'Z

Step 3: Perform Partition Design Experiment Planning is key Requires increased coordination between
process steps Identification of Outputs and Inputs
13
Run Order Std Order A B C D R1 R21 2 1 1 1 1 34.4 31.932 1 -1 1 -1 1 19.9 21.753 4 1 -1 1 1 4.7 11.724 5 -1 -1 1 -1 -31.3 1.8385 3 1 1 -1 -1 13.4 -27.64
Partition 2 ResponsesPartition 1

Step 4: Perform Partition Design AnalysisFor Each Response:
A. Select Significant Effects
B. Build Empirical Model
C. Calculate Partition Intercept Response
D. Select Significant Effects for Intercept Response
14

Step 4a: Select Significant Effects
15
Sum of Mean F p-valueSource Squares df Square Value Prob > FModel 2428.235 3 809.4116 27528.08 0.0044 A-A 246.0159 1 246.0159 8366.999 0.0070 B-B 1022.879 1 1022.879 34788.12 0.0034 D-D 515.6769 1 515.6769 17538.17 0.0048
Residual 0.029403 1 0.029403Cor Total 2428.264 4
R-Squared 0.999988Adj R-Squared 0.999952

Step 4a: Select Significant Effects
16
STEP 4a - First PartitionSum of Mean F p-value
Source Squares df Square Value Prob > FModel 1912.56 2 956.28 3.71 0.2124 A-A 365.77 1 365.77 1.42 0.3558 B-B 1266.76 1 1266.76 4.91 0.157
Residual 515.71 2 257.85Cor Total 2428.26 4
R-Squared 0.7876Adj R-Squared 0.5752

Step 4b: Build Empirical Model
17
Final Equation in Terms of Coded Factors:
R1 =3.158.85 * A
16.48 * B

Step 4c:Calculate Partition Intercept Response
18
A B C D R1 Int11 1 1 1 34.4 9.101-1 1 -1 1 19.9 12.3181 -1 1 1 4.7 12.317-1 -1 1 -1 -31.3 -6.0111 1 -1 -1 13.4 -11.959
Calculations
Int1i = - 8.85A - 16.47B + y1i
for i= 1 to N
Run 1
Int1i = - 8.85A - 16.47B + y1i
Int11 = - 8.85(1) - 16.47(1) + 34.4
Int11 = 9.101

Step 4: Partition Analysis
Repeat for Second PartitionA. Select Significant Effects
B. Build Empirical Model
C. Calculate Partition Intercept Response
19
A B C D R1 R2 Int1 Int21 1 1 1 34.4 31.93 9.101 9.298-1 1 -1 1 19.9 21.75 12.318 11.7941 -1 1 1 4.7 11.72 12.317 -10.912-1 -1 1 -1 -31.3 1.838 -6.011 11.7941 1 -1 -1 13.4 -27.64 -11.959 -5.008

Step 4d:Select Significant Effects for Intercept
20
Sum of Mean F p-valueSource Squares df Square Value Prob > FModel 491.1211 1 491.1211 59.92901 0.0045 AC 491.1211 1 491.1211 59.92901 0.0045
Residual 24.58514 3 8.195048Cor Total 515.7063 4
R-Squared 0.9523Adj R-Squared 0.9364

Step 5:Build Final Empirical Model
21
Final Equation in Terms of Coded Factors:
R1 = R2 =1.635938 1.9747.335938 A 4.919 C14.95844 B 14.875 D10.62094 AC 9.934 BD

ManufacturingProcess #1
Controllable factors
Uncontrollable factors
Inputs
Outputs, y
x1 x2 xk
z1 z2 zr
. . .
. . .
• Q8 Design Space– Link input parameters with quality attributes over broad range
• Traditional Design of Experiments (DOE)– Systematic approach to study effects of multiple factors on process
performance
– Limitation: not applied to multiple sequential process steps; does not account for the effects of upstream process factors on downstream process outputs
Case Study: Biogen IDEC
ManufacturingProcess #2
Controllable factors
Uncontrollable factors
Inputs
Outputs, y
x1 x2 xk
z1 z2 zr
. . .
. . .
22

Protein-A
Controllable factors
Uncontrollable factors
x 1 x 2 x k
z 1 z 2 z r
. . .
. . .
CIEX
Controllable factors
Uncontrollable factors
x 1 x 2 x k
z 1 z 2 z r
. . .
. . .
pH 4.5 poolpH 5.75 poolpH 7 pool
20 Protein-A eluate pools
20 CEXeluate pools
Harvest
Controllable factors
Uncontrollable factors
x 1 x 2 x k
z 1 z 2 z r
. . .
. . .
Case Study: Biogen IDECPartition Design: Experimental
Resolution IV: 1/16 fractional factorial for whole design
Each partition: full factorial
Harvest pH included in Protein A partition
Each column: 16 expts + 4 center points = 20 expts
23

Partition Design: Designs
ExperimentHarvest
pH
Load Capacity
(%)
Wash I Conc. (mM)
Elution velocity (cm/hr)
1 5.75 75 2100 262.52 4.5 120 0 753 7 30 0 754 7 30 0 4505 4.5 30 0 756 4.5 30 4200 757 7 30 4200 758 4.5 30 4200 4509 7 120 4200 75
10 5.75 75 2100 262.511 5.75 75 2100 262.512 4.5 30 0 45013 7 120 0 7514 4.5 120 0 45015 7 120 0 45016 4.5 120 4200 7517 7 30 4200 45018 4.5 120 4200 45019 7 120 4200 45020 5.75 75 2100 262.5
Protein-A Chromatography Step
Mab Eluate from
Experiment #
Load Capacity
(%)
Wash volume
(CV)
Elution NaCl Conc. (mM)
Elution pH
1 70 3 155 5.52 110 2 185 63 30 4 185 64 110 2 185 55 30 2 125 56 110 4 185 57 110 2 125 68 30 2 185 69 30 2 185 5
10 70 3 155 5.511 70 3 155 5.512 110 4 125 613 110 4 125 514 30 4 185 515 30 2 125 616 30 4 125 617 30 4 125 518 110 2 125 519 110 4 185 620 70 3 155 5.5
Cation Chromatography Step
24

25
Input Parameter% of Total
Sum of Squares
Load HCP [f(Harvest pH, ProA Wash I)]
83.3
CIEX Elution pH 6.6
Load HCP2 5.8
Load HCP * Elution pH 1.6
CIEX Elution [NaCl] 1.1
CIEX Elution pH2 0.7
CIEX Elution [NaCl] * CIEX Elution pH 0.5
Load HCP * CIEX Elution [NaCl] 0.2
CIEX Load Capacity 0.1
R2 0.96
Adjusted R2 0.95
Predicted R2 0.92
Input Parameter
% of Total Sum of Squares
Harvest pH 32.6
Pro A Wash I Conc. 17.7
Harvest pH * Pro A Wash I 15.6
CIEX Elution pH 10.1
Harvest pH * CIEX Elution pH 8.8
Pro A Wash I. * CIEX Elution pH 4.6
CIEX Load Capacity 3.9
Pro A Wash I. Conc. * CIEX Elution NaCl 1.8
CIEX Elution [NaCl] 1.5
Harvest pH * CIEX Elution [NaCl] 1.2
CIEX Elution [NaCl] * CIEX Elution pH 0.2
R2 0.99
Adjusted R2 0.99
Predicted R2 0.96
Traditional Model Results Partition Model Results
CIEX Step HCP ANOVA Comparison: Main Effects
• Partition model identified same significant main factors and their relative rank in significance

26
Input Parameter% Sum
of Squares
Load HCP [f(A,C)] 83.3
CIEX Elution pH 6.6
Load HCP2 5.8
Load HCP * CIEX Elution pH 1.6
Elution [NaCl] 1.1
Elution pH2 0.7
Elution [NaCl] * CIEX Elution pH 0.5
Load HCP * CIEX Elution [NaCl] 0.2
SPXL Load Capacity (mg/ml) 0.1
Input Parameter% of
Total Sum of Squares
Harvest pH 32.6
Pro A Wash I Conc. 17.8
Harvest pH * Pro A Wash I conc 15.6
CIEX Elution pH 10.1
Harvest pH * CIEX Elution pH 8.8
Pro A Wash I. Conc.* CIEX Elution pH 4.6
CIEX Load Capacity 3.9
ProA Wash 1. * CIEX Elution NaCl 1.8
Elution [NaCl] 1.5
Harvest pH * CIEX Elution [NaCl] 1.2
CIEX Elution [NaCl] * CIEX Elution pH
0.2
Traditional Model Results Partition Model Results
• Partition model able to identify interactions between process steps
CIEX Step HCP ANOVA Comparison: Interactions

27
Summary of Partition Designs
Experimental design capable of handling a serial process Sequential process steps that possess several factors and multiple
responses
Potential Advantages Links process steps together: identify upstream operation effects
and interactions to downstream processes. Better understanding of the overall process Potentially less experiments No manipulation of uncontrollable parameters necessary
ManufacturingProcess #1
Controllable factors
Uncontrollable factors
Inputs
Outputs, y
x1 x2 xk
z1 z2 zr
. . .
. . .
ManufacturingProcess #2
Controllable factors
Uncontrollable factors
Inputs
Outputs, y
x1 x2 xk
z1 z2 zr
. . .
. . .
ManufacturingProcess #3
Controllable factors
Uncontrollable factors
Inputs
Outputs, y
x1 x2 xk
z1 z2 zr
. . .
. . .

References
D. E. Coleman and D. C. Montgomery (1993), ‘Systematic Approach to Planning for a Designed Industrial Experiment’, Technometrics, 35, 1-27.
Lin, D.J.K. (1993). "Another Look at First-Order Saturated Designs: The p-efficient Designs," Technometrics, 35: (3), p284-292.
Montgomery, D.C., Borror, C.M. and Stanley, J.D., (1997). “Some Cautions in the Use of Plackett-Burman Designs,” Quality Engineering, 10, 371-381.
Box, G. E. P. and Draper, N. R. (1987) Empirical Model Building and Response Surfaces, John Wiley, New York, NY
Box, G. E. P. and Wilson, K. B. (1951), “On the Experimental Attainment of Optimal Conditions,” Journal of the Royal Statistical Society, 13, 1-45.
Hartley, H. O. (1959), “Smallest composite design for quadratic response surfaces,” Biometrics 15, 611-624.
Khuri, A. I. (1988), “A Measure of Rotatability for Response Surface Designs,” Technometrics, 30, 95-104.
Perry, L. A., Montgomery, and D. C, Fowler, J. W., " Partition Experimental Designs for Sequential Processes: Part I - First Order Models ", Quality and Reliability Engineering International, 18,1.
28