Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
\
|
=
|
|
.
|
\
|
=
|
|
.
|
\
|
=
=
=
n
i
i S
n
i
i
L
T
y
n
SN
y n
SN
S
y
SN
1
2
10
1
2
10
2
2
10
1
log 10
1 1
log 10
log 10
JMP will select the correct SN ratio when the factor goals are defined
Tends to be correlated to the
standard deviation
Correlates to the mean
Correlates to the mean
Better to optimize on the Mean and Stdev as separate Ys than to
use the SN ratio
158
TLC, DOE 10104
Taguchi Analysis Example
Open Taguchi Array L8.jmp for analysis. Examine the
design structure, the way the Y data is summarized:
mean, SN Ratio and standard deviation.
Mean
SN
L
Ratio
Stdev
159
TLC, DOE 10104
Taguchi Example As Is Analysis
Select Analyze, Fit Model.
Notice the noise factors are
not part of the analysis and
therefore their effects are
not determined (if any).
160
TLC, DOE 10104
Taguchi Example As Is Analysis
93
94
95
96
97
98
99
M
e
a
n
Y
i
e
l
d
A
c
t
u
a
l
93 94 95 96 97 98 99
Mean Yield Predicted P=0.1316 RSq=1. 00
RMSE=0.3536
Actual by Predicted Plot
RSquare
RSquare Adj
Root Mean Square Error
Mean of Response
Observations (or Sum Wgts)
0.995042
0.965293
0.353553
95.90625
8
Summary of Fit
Model
Error
C. Tot al
Source
6
1
7
DF
25. 085938
0. 125000
25. 210938
Sum of Squares
4. 18099
0. 12500
Mean Square
33. 4479
F Ratio
0. 1316
Prob > F
Analysis of Variance
39.5
39.6
39.7
39.8
39.9
S
N
R
a
t
i
o
Y
i
e
l
d
A
c
t
u
a
l
39.5 39.6 39.7 39.8 39.9
SN Rat io Yield Predicted P=0. 1321
RSq=1. 00 RMSE=0.0321
Actual by Predicted Plot
RSquare
RSquare Adj
Root Mean Square Error
Mean of Response
Observations (or Sum Wgts)
0.995001
0. 96501
0.032113
39.63389
8
Summary of Fit
Model
Error
C. Tot al
Source
6
1
7
DF
0.20528103
0.00103127
0.20631230
Sum of Squares
0.034214
0.001031
Mean Square
33. 1761
F Ratio
0. 1321
Prob > F
Analysis of Variance
Mean Yield SN
L
Yield
M
e
a
n
Y
ie
ld
102.327
89.9228
98.125
4.2022
S
N
R
a
t
io
Y
ie
l
d
40.2161
39.0915
39.83436
0.3817
D
e
s
ir
a
b
ili
t
y
1
0
0.641042
Temp
2
5
0
3
0
0
Pressure
1
5
3
0
Speed
1
0
0
2
0
0
Desirability
0 1
Prediction Profiler
Notice the SN ratio
and the mean are
highly correlated for
the SN
L
case.
161
TLC, DOE 10104
SN
L
, Mean and Standard Deviation Correlations
0.5
1.0
1.5
2.0
S
t
d
D
e
v
Y
i
e
l
d
39.5 39.6 39.7 39.8 39.9
SN Rat io Yield
Linear Fit
Fit Mean
St d Dev Yield = -19. 43921 + 0. 5190302 SN Ratio Yield
RSquare
RSquare Adj
Root Mean Square Error
Mean of Response
Observations (or Sum Wgt s)
0.036074
-0. 12458
0.497516
1.131976
8
Summary of Fit
Model
Error
C. Tot al
Source
1
6
7
DF
0. 0555790
1. 4851349
1. 5407138
Sum of Squares
0.055579
0.247522
Mean Square
0. 2245
F Ratio
0. 6524
Prob > F
Analysis of Variance
I nt ercept
SN Rat io Yield
Term
-19. 43921
0.5190302
Est imat e
43.41253
1.095329
St d Error
-0. 45
0. 47
t Rat io
0.6700
0.6524
Prob>| t|
Parameter Estimates
Linear Fit
Bivariate Fit of Std Dev Yield By SN Ratio Yield
0.5
1.0
1.5
2.0
S
t
d
D
e
v
Y
i
e
l
d
93 94 95 96 97 98 99
Mean Yield
Linear Fit
Fit Mean
St d Dev Yield = -3. 550964 + 0.0488283 Mean Yield
RSquare
RSquare Adj
Root Mean Square Error
Mean of Response
Observ at ions (or Sum Wgts)
0.039013
-0. 12115
0.496757
1.131976
8
Summary of Fit
Model
Error
C. Tot al
Source
1
6
7
DF
0. 0601080
1. 4806058
1. 5407138
Sum of Squares
0.060108
0.246768
Mean Square
0. 2436
F Ratio
0. 6392
Prob > F
Analysis of Variance
I nt ercept
Mean Yield
Term
-3. 550964
0.0488283
Est imat e
9.490102
0.098935
St d Error
-0. 37
0. 49
t Rat io
0.7211
0.6392
Prob>| t|
Parameter Estimates
Linear Fit
Bivariate Fit of Std Dev Yield By Mean Yield
93
94
95
96
97
98
99
M
e
a
n
Y
i
e
l
d
39.5 39.6 39.7 39.8 39.9
SN Rat io Yield
Linear Fit
Fit Mean
Mean Yield = -342. 2022 + 11. 053882 SN Ratio Yield
RSquare
RSquare Adj
Root Mean Square Error
Mean of Response
Observ at ions (or Sum Wgts)
0.999921
0.999908
0.018191
95.90625
8
Summary of Fit
Model
Error
C. Tot al
Source
1
6
7
DF
25. 208952
0. 001985
25. 210938
Sum of Squares
25. 2090
0. 0003
Mean Square
76181. 04
F Ratio
<. 0001
Prob > F
Analysis of Variance
I nt ercept
SN Rat io Yield
Term
-342. 2022
11.053882
Est imat e
1. 58731
0.040049
St d Error
-215. 6
276.01
t Rat io
<.0001
<.0001
Prob>| t|
Parameter Estimates
Linear Fit
Bivariate Fit of Mean Yield By SN Ratio Yield
Use Fit Y by X to examine the correlation of the Y terms
162
TLC, DOE 10104
Taguchi Example Improved Analysis
Open Taguchi L8 Modified.jmp.
Select Analyze, Fit Model and add
all factors and two factor
interactions using Macro, Factorial
to Degree.
163
TLC, DOE 10104
Predicting Yields with All Significant Terms
90
92
94
96
98
100
102
Y
i
e
l
d
A
c
t
u
a
l
92 94 96 98 100
Yield Predict ed P<. 0001 RSq=0.93
RMSE=0.6626
Actual by Predicted Plot
RSquare
RSquare Adj
Root Mean Square Error
Mean of Response
Observations (or Sum Wgts)
0.932324
0.900097
0.662562
95.90625
32
Summary of Fit
Model
Error
C. Tot al
Source
10
21
31
DF
127. 00000
9.21875
136. 21875
Sum of Squares
12. 7000
0. 4390
Mean Square
28. 9302
F Ratio
<. 0001
Prob > F
Analysis of Variance
Temp
Pressure
Speed
Operator
Machine
Temp*Speed
Temp*Operator
Temp*Machine
Pressure*Speed
Operator*Machine
Source
1
1
1
1
1
1
1
1
1
1
Nparm
1
1
1
1
1
1
1
1
1
1
DF
94.531250
0.500000
0.781250
16.531250
2.531250
2.531250
2.531250
2.531250
2.000000
2.531250
Sum of Squares
215.3390
1.1390
1.7797
37.6576
5.7661
5.7661
5.7661
5.7661
4.5559
5.7661
F Ratio
<.0001
0.2980
0.1965
<.0001
0.0257
0.0257
0.0257
0.0257
0.0448
0.0257
Prob > F
Effect Tests
Y
i
e
l
d
100.214
92
99.40625
0.8078
D
e
s
i
r
a
b
i
l
i
t
y 1
0
0.909895
Temp
2
5
0
3
0
0
250
Pressure
1
5
3
0
15
Speed
1
0
0
2
0
0
200
Operator
F
r
e
d
M
a
r
y
Machine
M
1
M
2
Desirability
0 1
Prediction Profiler
All of the noise is correctly accounted for in this example
164
TLC, DOE 10104
Limitations of Taguchi Designs
Interactions are active but often aliased
Model interpolation is not exploited to find better settings than
actually tested
SN ineffective in identifying dispersion effects
Formula confounds location and dispersion effects, not a very
good idea
With the bad news comes the good news, Taguchi experiments
are far better than no experimentation and are founded on
technically correct full and fractional factorial designs, the major
issues with Taguchi are less on design and more on analysis.
An improvement to Taguchi style signal to noise modeling can
be done by using the mean and standard deviation of the
replicates as individual Ys
165
TLC, DOE 10104
Section VI
Introduction to DOE and robust design principles
Experimental preparation
Full factorial designs
Screening designs
Taguchi designs (optional)
Custom designs
Optimization designs
Mixture designs (optional)
166
TLC, DOE 10104
Characteristics of Custom Designs
Custom designs
Have the most flexibility of all design choices
Can be used in situations not suitable for classic designs
Can be used for routine factor screening, response optimization,
and mixture problems
167
TLC, DOE 10104
Generality of Custom Designs
Custom designs can be generated for
Categorical factors with any number of levels
Continuous factors (any number)
Blocks with user-specified number of runs
Mixture ingredients
Covariates (fixed)
Inequality constraints on factors
User-specified models
Models where the number of runs is greater than or equal the
number of terms in the model
Models where the number of runs is less than the number of terms
in the model
168
TLC, DOE 10104
Custom Design Generation
The advantages:
Does not use factorial combinations avoids large and inefficient
designs
Does not require a candidate set of design points avoids long
iterations
Provides a flexible approach to most problems
Enables the user to specify the factors, model, and size
You cannot expect every real problem to follow the simple pattern
portrayed in book examples
The simple, tidy sequence of screen-optimize-confirm may restrict
the path that is dictated by your problem or directed by your data
and early findings
Experimental stages are often not clearly delineated; there usually
is some overlap
169
TLC, DOE 10104
Custom Design Benefits
The benefits include:
Any kind of factor: continuous, categorical, mixture, blocking,
covariate, constant
Any number of levels
Any number, combination of factors, levels
Any combination of effects in model
Any number of runs
Irregular experimental region
Treats the engineer or scientist as a customer when designing the
experiment
170
TLC, DOE 10104
D-optimal and I-optimal Designs
Open demoCoordinateExchange.jsl script to visualize the difference
between D-optimal and I-optimal designs. This script also
demonstrates the general coordinate exchange approached used by
JMP in generating the experimental matrix for custom designs.
Developed by Bradley Jones, SAS
D-Optimal designs are generated
when two level and interaction
designs are requested. D-Optimal
designs emphasize the corners of
the design space and improve
coefficient estimation.
I-Optimal designs are generated
when RSM designs are requested.
I-Optimal designs emphasize the
center of the design space and
improve the variance estimation
across the range of the experiment.
171
TLC, DOE 10104
Select DOE,
Custom
Design from
the JMP
Starter Menu
Custom Designs in JMP
172
TLC, DOE 10104
Add, the responses, Goals
and Limits, then add the
number of factors you need for
the experiment and define the
Name and Values (levels)
then select Continue
Custom Designs in JMP
Y response importance is
evenly weighted unless
specified. Importance is a
relative weighting factor.
The values for the Xs are Max
and Min values only level
setting occurs in the model
specification.
Notice the range of factor
types that can be used in the
experiment.
173
TLC, DOE 10104
Select the factors to be
used in the experiment then
select RSM for response
surface model.
This will give you a list of all
the linear terms, the
quadratic terms, and all two
factor interaction terms.
You can then add or
subtract terms that you
consider appropriate and
you will see the impact on
the sample size at the
bottom.
Custom Designs in JMP
174
TLC, DOE 10104
Understanding Custom Design Options
Main Effects
Linear terms in the model
Interactions
Any or all two factor+ interactions
RSM (Response Surface
Methodology)
All main effects, two factor
interactions and quadratic terms
Cross
Any combination of power and
interaction terms
Power (2=3 levels etc.)
Any level of a continuous X variable,
supports 3-6 levels
175
TLC, DOE 10104
Design Generation
Minimum
Generates the smallest DOE
possible with all effects in the
model estimable and no df
allocated for error (saturated design)
Default
Full factorial design increments the model terms fit within (4, 8, 16, 32 etc)
Compromise
Compromise is a second suggestion that is more conservative than the
Default. Its value is generally between Default and Grid.
Grid
Full factorial design, for this example 3
4
User Specified
Any number of runs above the minimum design, a good rule of thumb is to
make sure the design has at least two runs over minimum.
176
TLC, DOE 10104
Select User
Defined 17, Make
Design. The next
step is to evaluate
the design and
make sure it has the
properties you want.
Custom Designs in JMP
177
TLC, DOE 10104
Variance Prediction Profiler
Prediction Variance Profiler
is used to determine the
relative variability in the
response depending on the
DOE design. Add additional
samples and the variance
goes down. Values of greater
than 1 are undesirable.
Modify the design to 20 runs
and see what happens to the
predicted variance. Is best
used to compare alternative
designs. Notice the variance
is lowest at the center and
higher at the tails (I-Optimal
design).
178
TLC, DOE 10104
Prediction Variance Surface
Provides a more complete view of the effect of sample size on the
predicted variance of the DOE structure. More points in the corners will
cause the predicted variance to go down.
179
TLC, DOE 10104
Now lets
evaluate the
correlation of
the Xs to see
if this is an
acceptable
design.
Select
Analyze,
Multivariate
Methods,
Multivariate
Evaluation of Custom Designs in JMP
180
TLC, DOE 10104
Looks good no r values
above .1 or 10%
The custom design tool
allowed for more information
in less number or runs. This
design allows for the
estimation of all linear
effects, two factor
interactions and curvature.
Open Extrusion
Experiment RSM.jmp and
complete the analysis of a
this 16 run custom design.
Evaluation of Custom Designs
181
TLC, DOE 10104
Analysis of Custom Design
O
u
t
p
u
t
1638
812
1376.795
18.695
D
i
a
m
e
t
e
r
2.661
2.38294
2.540232
0.0234
C
r
a
c
k
s
17.7
1.00719
3.754238
0.9347
D
e
s
i
r
a
b
i
l
i
t
y
1
0
0.80278
Speed
1
0
0
2
0
0
162.211
Temp
2
5
0
3
0
0
289.385
Time
5
1
0
5
Pressure
1
5
3
0
25.8441
Desirability
0 1
Prediction Profiler
182
TLC, DOE 10104
Interaction Plots
800
1200
1600
O
u
t
p
u
t
800
1200
1600
O
u
t
p
u
t
800
1200
1600
O
u
t
p
u
t
800
1200
1600
O
u
t
p
u
t
Speed
250
300
5
10
15
30
100 150 200
100
200
Temp
5
10
15
30
250 270 290 310
100
200
250
300
Time
15 30
5 6 7 8 9 10 11
100
200
250
300
5
10
Pressure
15 20 25 30
S
p
e
e
d
T
e
m
p
T
i
m
e
P
r
e
s
s
u
r
e
2.35
2.45
2.55
2.65
D
i
a
m
e
t
e
r
2.35
2.45
2.55
2.65
D
i
a
m
e
t
e
r
2.35
2.45
2.55
2.65
D
i
a
m
e
t
e
r
2.35
2.45
2.55
2.65
D
i
a
m
e
t
e
r
Speed
250
300
5
10
15
30
100 150 200
100
200
Temp
5
10
15
30
250 270 290 310
100
200
250
300
Time
15
30
5 6 7 8 9 10 11
100
200
250
300
5
10
Pressure
15 20 25 30
S
p
e
e
d
T
e
m
p
T
i
m
e
P
r
e
s
s
u
r
e
0
5
10
15
20
C
r
a
c
k
s
0
5
10
15
20
C
r
a
c
k
s
0
5
10
15
20
C
r
a
c
k
s
0
5
10
15
20
C
r
a
c
k
s
Speed
250
300
5
10
15
30
100 150 200
100
200
Temp
5
10
15
30
250 270 290 310
100
200
250
300
Time
15
30
5 6 7 8 9 10 11
100
200
250
300
5
10
Pressure
15 20 25 30
S
p
e
e
d
T
e
m
p
T
i m
e
P
r
e
s
s
u
r
e
Select
Interaction
Profiler to see
all two factor
interactions
183
TLC, DOE 10104
Surface Plots
Use Surface
Plots for
visualization and
presentation of
model effects
and sensitivities
184
TLC, DOE 10104
Supersaturated Designs
Supersaturated designs are those
designs that have more model effects
than the number of experimental runs.
Necessary
Must be in the design and must be
estimable
If Possible
Are estimable as long as the number of
significant terms in the model are less than
the experimental sample size. The larger
the experiment the more probable
experimental terms will not be significant.
Open Custom Design and select 16
factors, make four of them Necessary and
the balance of them If Possible. Set User
Specified to 14 runs. Notice minimum is
now 5.
185
TLC, DOE 10104
Supersaturated Designs
Although the correlation of the X variables are generally higher than
desirable (0 is the goal) it does not prevent us from running the
experiment and analyzing the response variables. Maximum
correlation of the factors is .4 or r
2
=16%
186
TLC, DOE 10104
The Pareto Principle
Based on the Pareto
principle we do not
expect all factors to
have an effect. This
is particularly true
when running
screening
experiments.
Simulate the
following responses,
then use forward
stepwise to find the
effects, remove all
effects less than one
and then fit the
model and examine
the Scaled
Estimates.
187
TLC, DOE 10104
Design Tips for Blocking in Custom Designs
There are two approaches for specification of blocking in JMP
Method 1 Block Size Unknown
1. Define all factors and responses
2. Determine the size of the experiment with all model terms
3. Determine the blocking variable and the sample size for each block
then tell JMP the block size for the smallest group and label the
blocks accordingly
Method 2 Known Block Size
1. Define all factors and responses
2. Determine the blocking variable and the sample size for each block
3. Determine the size of the experimental and all model terms
4. JMP will correctly define the number of blocks required, then label
each block
188
TLC, DOE 10104
Blocking Exercise
Design an experiment for the following conditions, use a Custom
Design
Exercise: Method 2 Block Size Known:
5 Continuous Factors
Parts from the experiment come from multiple lots with 6 parts to the lot.
You are worried the lot effect could influence the response. You want to
model all linear terms only. Assuming you block for the lot, design the
experiment. What would be size of the experiment if you ignored the lot
effect?
In Class Exercise: 5 Min.
189
TLC, DOE 10104
Blocking Analysis
Open Blocked for Tester.jmp
190
TLC, DOE 10104
Blocking Analysis
Conclusion, blocking for tester
isolated a 4.5% yield effect.
If the F value for the blocked
variable is >1 then blocking
improved the experiment if F is
<1 then there was no value in
blocking for the factor of
concern.
I nt ercept
Power(150,210)
Temperature(250,275)
Pressure(750, 1000)
Power*Power
Power*Temperature
Temperature*Temperature
Power*Pressure
Temperature*Pressure
Pressure*Pressure
Tester[1]
Term
82. 53125
10. 05
0.55
-0.25
1. 46875
0.3125
-12. 53125
-1.0625
-0.6875
-0. 03125
2.25
Est imat e
0.917005
0.427639
0.427639
0.427639
0.917005
0.478115
0.917005
0.478115
0.478115
0.917005
0.361421
St d Error
90. 00
23. 50
1. 29
-0. 58
1. 60
0. 65
-13. 67
-2. 22
-1. 44
-0. 03
6. 23
t Rat io
<.0001
<.0001
0.2547
0.5842
0.1701
0.5422
<.0001
0.0769
0.2100
0.9741
0.0016
Prob>| t|
Parameter Estimates
Power(150,210)
Temperature(250,275)
Pressure(750, 1000)
Power*Power
Power*Temperature
Temperature*Temperature
Power*Pressure
Temperature*Pressure
Pressure*Pressure
Tester
Source
1
1
1
1
1
1
1
1
1
1
Nparm
1
1
1
1
1
1
1
1
1
1
DF
1154. 3143
3. 4571
0. 7143
5. 3617
0. 8929
390.2937
10. 3214
4. 3214
0. 0024
81. 0000
Sum of Squares
552.3035
1. 6541
0. 3418
2. 5654
0. 4272
186.7434
4. 9385
2. 0677
0. 0012
38. 7560
F Ratio
<. 0001
0. 2547
0. 5842
0. 1701
0. 5422
<. 0001
0. 0769
0. 2100
0. 9741
0. 0016
Prob > F
Effect Tests
Y
i
e
l
d
100.928
59
97.50958
3.1259
D
e
s
i
r
a
b
i
l
i
t
y 1
0
0.938989
Power
1
5
0
2
1
0
210
Temperature
2
5
0
2
7
5
263.123
Pressure
7
5
0
1
0
0
0
761.439
Tester
1 2
Desirability
0 1
Prediction Profiler
191
TLC, DOE 10104
Fixed Covariates as Part of the Experiment
1. Design experiment without covariate and
determine number of runs required +1 for the
covariate + n for any additional model terms
associated with the covariate
2. Measure the materials or parts prior to
inclusion in the experiment. Place the data
in a column with the correct number of
samples measured.
3. Add the covariate to the experiment and
finish the design
4. If more than one covariate needs to be
added, add the one that you believe is the
most likely to influence the Ys
Open Material Density.jmp
192
TLC, DOE 10104
Fixed Covariates
Design the experiment and
determine the number of runs
required.
Then pull the factor and data into
the experiment.
JMP will make the covariate
orthogonal to the rest of the
experiment.
193
TLC, DOE 10104
Use Multivariate to Examine the Matrix
Notice JMP using the
coordinate exchange
method was able to
position the covariate as
to minimize the correlation
of the covariate to other
factors
194
TLC, DOE 10104
Fix Covariate Analysis
Open and analyze Factory
Simulation with Fixed
Covariate.jmp
O
u
t
p
u
t
1649
779
1205.107
20.52
D
i
a
m
e
t
e
r
2.665
2.381
2.528952
0.0034
C
r
a
c
k
s
19.1
2.9
7.882989
0.3563
Speed
1
0
0
2
0
0
150
Temperature
2
5
0
3
0
0
275
Time
5
1
0
7.5
Pressure
1
5
3
0
22.5
Mat erial
Densit y
2
0
2
9
24.5
Prediction Profiler
195
TLC, DOE 10104
Covariates During the Experiment
Measure the uncontrolled
covariate during the
experiment. Examples
may be temperature,
pressure, time, ph, force,
ESD, etc.
Notice under Method:
REML appears. (residual
maximum likelihood)
Add the covariate column
to the JMP table and add
the covariate to the
analysis when fitting a
model. Use the red
attribute triangle and set
the covariate to random.
Bounce Data with Covariate.JMP
196
TLC, DOE 10104
Insitu Covariate Analysis
Examining the REML
analysis shows ~23% of
the unexplained variation
(noise) can be explained
as due to temperature
variation. Further
examination shows the
effect of temperature to
be non significant
(p>.05). Temperature is
not an important
variable across the
range examined in the
experiment.
Temperature&Random
Residual
Total
Random Ef fect
0.3025457
Var Ratio
0.9293855
3.0718852
4.0012708
Var Component
2.4976207
Std Error
0.1062835
95% Lower
7.6557e10
95% Upper
23.227
76.773
100.000
Pct of Total
-2 LogLikelihood = 37.729433
REML Variance Component Estimates
Continuous f actors centered by mean, scaled by range/2
Intercept
Silica(0.7,1.7)&RS
Silane(40,60)&RS
Sulf ur(1.8,2.8)&RS
Silane*Silica
Sulf ur*Silica
Sulf ur*Silane
Silica*Silica
Silane*Silane
Sulf ur*Sulfur
Temperature&Random
Term
396.35001
-40.17122
10.694971
-26.52375
0.2121657
114.68487
46.397893
26.225782
-0.172111
-35.45995
-1.109198
Scaled Estimate
1.053412
0.62087
0.670137
0.685788
0.957818
0.889878
0.885028
0.940521
0.978076
0.912738
0.927792
Std Error
376.25
-64.70
15.96
-38.68
0.22
128.88
52.43
27.88
-0.18
-38.85
-1.20
t Ratio
<.0001
<.0001
<.0001
<.0001
0.8355
<.0001
<.0001
<.0001
0.8689
<.0001
0.2979
Prob>| t|
Scaled Estimates
S
t
r
e
t
c
h
570
260
396.35
2.9247
Silica
0
.
7
1
.
7
1.2
Silane
4
0
6
0
50
Sulf ur
1
.
8
2
.
8
2.3
Temperature
1
3
9
1
4
2
140.593
Prediction Profiler
197
TLC, DOE 10104
Mixture Factors
Open the Etch Mixture.JMP file
and analyze the data. Note how
the profiler keeps the mixtures
organized during the analysis and
modeling.
198
TLC, DOE 10104
Mixture Analysis
Notice the mixture relationship
is still maintained
199
TLC, DOE 10104
Setting Experimental Constraints in JMP
Factor constraints are set in the following way:
1. Determine the levels of the factor settings in the
experimental units without respect to a constraint.
2. Determine the linear multiplier for each of the
factor setting (1=1*X, 2=2*X, -1=-1*X).
3. If a constraint is set to zero for a factor, JMP will
ignore its settings and set no constraint.
4. Setting a constraint to one for a factor will include
its setting with no additional linear scaling and set
a constraint per the conditions indicated at the
bottom.
5. After the multiplier has been defined, determine the
boundaries of the constraint by specifying less than
or equal or greater than or equal to a specified
value.
6. The user-defined value at the bottom of the
constraint is the sum of the constraints times the
factor settings (y=X1*C1+X2*C2 Xn*Cn).
7. After Make Design is selected, JMP will apply the
constraint to the generation of the DOE matrix.
200
TLC, DOE 10104
DOE Exercise Designing the Experiment
Using the factors and levels from your
brainstorming session, design the appropriate DOE
to characterize the process or product.
1) Use JMP to design your experiment
2) Use the custom design generator to create your DOE
matrix
3) Be prepared to share your results
In Class Exercise: 20 Min.
201
TLC, DOE 10104
Section I
Design and Analysis of Experiments (DOE)
Content
Introduction to DOE and robust design principles
Experimental preparation
Full factorial designs
Screening designs
Taguchi designs (optional)
Custom designs
Optimization designs
Mixture designs (optional)
Tools
JMP
202
TLC, DOE 10104
Section VII
Introduction to DOE and robust design principles
Experimental preparation
Full factorial designs
Screening designs
Taguchi designs (optional)
Custom designs
Optimization designs
Mixture designs (optional)
203
TLC, DOE 10104
Optimization Strategy
Once characterization has been done we are now ready to
consider optimization experiments when the data indicates
there is more improvement to be made
For factors which show little to no interaction they can be
optimized one factor at a time using a Oneway five level
design, or the Path of Steepest Assent method can be used
for two factors.
When factors interact they must be considered simultaneously
using a two factor Central Composite Design (CCD) or Box
Bhenken type design (5 level).
Local optimums are determined this way
204
TLC, DOE 10104
A design is rotatable if the variance of the predicted values is
the same at equal distances from the center of the design. By
adding the additional points the design is 5 level.
Properties of a CCD
Full
Factorial
Design
Rotatable
CCD RSM
Design
205
TLC, DOE 10104
Orthogonal
Center point
Orthogonal & Rotatable
Star or
Axial
points
Factorial
Points
Face
Centered
points
Options in a CCD Design
206
TLC, DOE 10104
2 Factor CCD
207
TLC, DOE 10104
3 Factor Box-Bhenken
3 Factor CCD
Box Bhenken versus CCD
Design both experiments using the Response Surface feature in
the DOE menu. Use the spin plot graph to examine the
experimental space.
208
TLC, DOE 10104
Yield Example
3 Factor CCD Yield.jmp
Follow our analysis
procedure and fit an
appropriate model.
209
TLC, DOE 10104
Using the Contour Profiler
Once we have
the model we can
use the Profiler
and Contour
Profiler to locate
the local
optimum. This
method makes it
easy to see
changes in the
response surface.
210
TLC, DOE 10104
Path of steepest assent uses
a 2x2 orthogonal design to
compute a vector pointing at
the steepest slope possible
from the data. Once the line
is known steps are taken to
move to the optimum.
Once at the local optimum a
CCD can be used to
determine if there is still
benefit to moving in another
direction.
Path of Steepest Assent
211
TLC, DOE 10104
Linear Model moving up the slope
y=|
o
+|
1
X
1
+|
2
X
2
+e
Quadratic Model near the top
y=|
o
+ |
1
X
1
+|
2
X
2
+ |
11
X
1
+|
22
X
2
+e
2
2
^
^
Two Model Equations
212
TLC, DOE 10104
2x2 full factorials are used to
determine the first order
linear model and steps are
taken until the response fails
to show improvement.
A new first order model is
then developed and the path
of steepest ascent is followed
until the first order model no
longer fits.
2X2 Design to Begin
213
TLC, DOE 10104
Response Process Yield
Factors Reaction Time (A) 30 - 40 Minutes
Temperature (B) 150 -160F
2x2 Full Factorial
Factor A Factor B Response
Time Temperature Yield
30 150 39.3
30 160 40.0
40 150 40.9
40 160 41.5
Procedure and Example
Path of Steepest Yield.jmp
214
TLC, DOE 10104
Terms of the Model:
y=|
o
+|
1
X
1
+|
2
X
2
+e
y=40.425+0.775X
1
+.325X
2
+0.05
X
i
=(X-Midpoint/1/2 Range) or (+1 to -1)
y=40.425+0.775((X-35)/5)+.325((X-155)/5)+0.05
^
^
^
Compute the Model
215
TLC, DOE 10104
y=40.425+0.775((X
1
-35)/5)+.325((X
2
-155)/5)+0.05
1. Determine a step size in one of the process variables say A X
1
.
Typically the factor which determines the step size is the one we know
the most about and or the largest coefficient. Begin with a coded step
size of 1 for A X
1
.
2. The coded step size for the second factor is a ratio of the first if we
wish to move up the slope in the most efficient manner.
3. Convert the coded steps into the factor steps by doing the following:
A Reaction time = AX
1
*1/2 original experiment A =1*5= 5
A Temperature = AX
2
*1/2 original experiment A =.42*5= 2
A
A
X
X
2
2
1 1
0 325
0 775 1
0 42 = = =
|
| ( / )
.
( . / )
.
Determining the Step Size
216
TLC, DOE 10104
Steps Factor Settings Response
Time Temp. Yield
Origin 35 155 GO
1 40 157 41.0
2 45 159 42.9
3 50 161 47.1
4 55 163 49.7
5 60 165 53.8
6 65 167 59.9
7 70 169 65.0
8 75 171 70.4
9 80 173 77.6
10 85 175 80.3
11 90 177 76.2
12 95 179 75.1
13 100 181 73.9
STOP
Move Up the Slope
Three in a row lower or higher than the optimal
setting indicates a local optimum
217
TLC, DOE 10104
0
10
20
30
40
50
60
70
80
90
0 1 2 3 4 5 6 7 8 9 10 11 12
Steps
Y
i
e
l
d
Local Optimum
Looking for the Local Optimum
13
218
TLC, DOE 10104
Section VIII
Introduction to DOE and robust design principles
Experimental preparation
Full factorial designs
Screening designs
Taguchi designs (optional)
Custom designs
Optimization designs
Mixture designs (optional)
219
TLC, DOE 10104
Factorial Designs
Each factor is free to vary independently of
all other factors
Experiments are arranged in square
orthogonal designs
Common designs are Full Factorial,
Fractional Factorial and D-optimal designs
RSM designs are used to characterize linear,
interaction and quadratic effects. Common
Designs are Box-Behnken and Central
Composite
220
TLC, DOE 10104
Each factor is not independent
Factors are a combination of fluids, gases
or ingredients which make up unity
Experiments are primarily Simplex or
triangular
Regression modeling still applies to the
analysis of mixture designs
Formulation of solids, fluids and gases in a
constant volume require mixture designs
Mixture Designs
221
TLC, DOE 10104
Mixture Designs
With mixtures it is impossible to vary one factor
independently of all the others. When you
change the amount of one ingredient, the
proportion of one or more of the other ingredients
must also change to make up the total.
This simple fact has a profound effect on every
aspect of experimentation with mixtures: the
factor space, the design properties, model and
the interpretation of the results.
222
TLC, DOE 10104
Two Factor Mixture Experiments and Pure Blends
Experimental points fall
along a straight line
The extremes of the line
are the Pure Blends of
the mixture
X
1
and X
2
are correlated
0
1
X
1
1
X
2
X
1
+ X
2
= 1
223
TLC, DOE 10104
Two Factor Mixture Experiments
You are not required to
experiment only with pure blends.
Add additional levels to determine
optimal performance.
The ratio of the two Xs
becomes the independent
factor in the experiment
0
1
X
1
1
X
2
X
1
+ X
2
= 1
224
TLC, DOE 10104
If Only Two Mixture Factors Treat Them as One
When conducting mixture experiments with only two factors treat them
as if they are a single factor
Gas 1 Gas 2 Ratio
1 75 25 3.0
2 50 50 1.0
3 25 75 0.3
4 25 75 0.3
5 75 25 3.0
6 25 75 0.3
7 75 25 3.0
8 50 50 1.0
9 50 50 1.0
225
TLC, DOE 10104
Simplex Designs
Because the proportions sum to
one, the feasible region for a
mixture takes the form of a
simplex.
Consider three factors in a 3
dimensional plot. The plane where
the sum of the three factors add to
one is a triangle-shaped slice.
0
1
1
1
X
2
X
1
X
3
X
1
+ X
2
+ X
3
= 1
226
TLC, DOE 10104
Mixture Designs in JMP
Begin designing
mixture
experiments by
selecting Mixture
Design from the
DOE menu.
227
TLC, DOE 10104
Simplex Lattice Designs
.2
.2
.2
.4
.4
.4
.6
.6
.6
.8
.8
.8
X2
X1
X3
228
TLC, DOE 10104
Definition of the Simplex Lattice Design
The simplex lattice design is a space-filling design that creates
a triangular grid of experiments
Uses (q,m) designation to define the simplex lattice design.
Where q is the number of factors and m is the number of levels.
Consists of points defined by the following coordinate settings:
the proportions assumed by each component take the m+1
equally spaced values from 0 to 1, X
i
= 0/m, 1/m, 2/m, , q/m
All possible combinations (mixtures) of the proportions are
created from this equation
229
TLC, DOE 10104
Simplex Lattice Design Example
Simplex lattice design (3,3)
3 factors at three levels + zero
0/m, 1/m, 2/m, 3/m
0/3 = 0, 1/3 = .33, 2/3 = .67, 3/3 = 1
X1 X2 X3
0 0 1
0 0.33 0.67
0 0.67 0.33
0 1 0
0.33 0 0.67
0.33 0.33 0.33
0.33 0.67 0
0.67 0 0.33
0.67 0.33 0
1 0 0
230
TLC, DOE 10104
Fitting Capability
.2
.2
.2
.4
.4
.4
.6
.6
.6
.8
.8
.8
X2
X1
X3
.2
.2
.2
.4
.4
.4
.6
.6
.6
.8
.8
.8
X2
X1
X3
.2
.2
.2
.4
.4
.4
.6
.6
.6
.8
.8
.8
X2
X3
X1
Linear Quadratic Cubic
(3,1) (3,2) (3,3)
one level two levels three levels
Depending on the number of levels Simplex Lattice designs support
the characterization of linear effects and higher order polynomials
231
TLC, DOE 10104
Number of Experiments for Lattice Designs
Number of Levels (m)
Number (2) (3) (4)
of Factors (q) Quadratic Cubic Quartic
(3) 6 10 15
(4) 10 20 35
(5) 15 35 70
(6) 21 56 126
(q+m-1)!
m!(q-1)!
232
TLC, DOE 10104
Simplex Lattice: (3,2) Design
.2
.2
.2
.4
.4
.4
.6
.6
.6
.8
.8
.8
X2
X3
X1
(3,2)
6 Experiment Matrix
Pure Blends
Binary Blends
X1 X2 X3
0 0 1
0 0.5 0.5
0 1 0
0.5 0 0.5
0.5 0.5 0
1 0 0
233
TLC, DOE 10104
(3,3)
10 Experiment Matrix
Simplex Lattice: (3,3) Design
. 2
. 2
. 2
. 4
. 4
. 4
. 6
. 6
. 6
. 8
. 8
. 8
X2
X1
X3
X1 X2 X3
0 0 1
0 0.33 0.67
0 0.67 0.33
0 1 0
0.33 0 0.67
0.33 0.33 0.33
0.33 0.67 0
0.67 0 0.33
0.67 0.33 0
1 0 0
234
TLC, DOE 10104
Simplex Lattice: (3,4) Design
(3,4)
15 Experiment Matrix
. 2
. 2
. 2
. 4
. 4
. 4
. 6
. 6
. 6
. 8
. 8
. 8
X2
X1
X3
X1 X2 X3
0 0 1
0 0.25 0.75
0 0.5 0.5
0 0.75 0.25
0 1 0
0.25 0 0.75
0.25 0.25 0.5
0.25 0.5 0.25
0.25 0.75 0
0.5 0 0.5
0.5 0.25 0.25
0.5 0.5 0
0.75 0 0.25
0.75 0.25 0
1 0 0
235
TLC, DOE 10104
Simplex Lattice: (3,5) Design
(3,5)
21 Experiment Matrix
Open and analyze Simplex Lattice.jmp
. 2
. 2
. 2
. 4
. 4
. 4
. 6
. 6
. 6
. 8
. 8
. 8
X2
X1
X3
X1 X2 X3
0 0 1
0 0.2 0.8
0 0.4 0.6
0 0.6 0.4
0 0.8 0.2
0 1 0
0.2 0 0.8
0.2 0.2 0.6
0.2 0.4 0.4
0.2 0.6 0.2
0.2 0.8 0
0.4 0 0.6
0.4 0.2 0.4
0.4 0.4 0.2
0.4 0.6 0
0.6 0 0.4
0.6 0.2 0.2
0.6 0.4 0
0.8 0 0.2
0.8 0.2 0
1 0 0
236
TLC, DOE 10104
Fitting Mixture Designs
When fitting a model for
mixture designs, you must take
into account that all the factors
add up to a constant, and thus
a traditional full linear model
will not be fully estimable
Suppress the intercept
Include all the linear main-
effect terms
Exclude all the squared terms
(X1*X1)
Include all the cross terms
(X1*X2)
Coefficients on the cross terms
indicate the curvature across
each edge of the factor space
237
TLC, DOE 10104
Simplex Centroid Designs
238
TLC, DOE 10104
Definition of Simplex Centroid Designs
A Simplex Centroid design of degree K with nf factors
K is the number of equal blends of the mixture components in
addition to one
+ a center point
All one factor
All combinations of two factors at equal levels
All combinations of three factors at equal levels and so on up
to K factors at a time combined at K equal levels
A center point run with equal amounts of all the ingredients is
always included
239
TLC, DOE 10104
Simplex Centroid Designs nf=3, K=1
X1 X2 X3
0 0 1
0 1 0
0.33 0.33 0.33
1 0 0
240
TLC, DOE 10104
Simplex Centroid Designs nf=3, K=2
X1 X2 X3
0 0 1
0 0.5 0.5
0 1 0
0.33 0.33 0.33
0.5 0 0.5
0.5 0.5 0
1 0 0
241
TLC, DOE 10104
Simplex Centroid Designs nf=3, K=3
X1 X2 X3
0 0 1
0 0.5 0.5
0 1 0
0.33 0.33 0.33
0.33 0.33 0.33
0.5 0 0.5
0.5 0.5 0
1 0 0
242
TLC, DOE 10104
Simplex Centroid Designs nf=4, K=4
X1 X2 X3 X4
0 0 0 1
0 0 0.5 0.5
0 0 1 0
0 0.33 0.33 0.33
0 0.5 0 0.5
0 0.5 0.5 0
0 1 0 0
0.25 0.25 0.25 0.25
0.25 0.25 0.25 0.25
0.33 0 0.33 0.33
0.33 0.33 0 0.33
0.33 0.33 0.33 0
0.5 0 0 0.5
0.5 0 0.5 0
0.5 0.5 0 0
1 0 0 0
Open and
analyze
Simplex
Centroid.jmp
243
TLC, DOE 10104
Extreme Vertices Designs
.2
.2
.2
.4
.4
.4
.6
.6
.6
.8
.8
.8
X2
X1
X3
Ternary Plot
244
TLC, DOE 10104
Want Something other than a Grid?
The extreme vertices design finds the corners (vertices) of a
factor space constrained by limits specified by the
experimenter for one or more of the factors
The property that the factors must be non-negative and must
add up to 1, is the basic mixture constraint that makes a
triangular-shaped region
Extreme Vertices are more flexible for the designs you want
245
TLC, DOE 10104
Definition of Extreme Vertices Designs
.2
.2
.2
.4
.4
.4
.6
.6
.6
.8
.8
.8
X2
X1
X3
Ternary Plot
Define the region as a set of vectors and
vertices across the design region.
Additional limits are usually in the form of
range constraints, upper bounds, and
lower bounds on the factor values.
Find the vertices of the simplex
Add the midpoints of the edges
Add the averages of the vertices to the
specified degree
246
TLC, DOE 10104
Extreme Vertices Only, No Constraint, Degree 2
X
1
>.1
X
1
<.5
X
2
>.1
X
2
<.7
X
3
>0
X
3
<.7
X
2
<.7
X
1
<.5
X
1
>.1
X
3
>0
X
3
<.7
X
2
>.1
247
TLC, DOE 10104
Linear Constraints
248
TLC, DOE 10104
Adding Linear Constraints
Sometimes other ingredients need
range constraints that confine their
values to be greater than a lower bound
or less than an upper bound.
Range constraints chop off parts of the
triangular-shaped (simplex) region to
make additional vertices. It is also
possible to have a linear constraint,
which defines a linear combination of
factors to be greater or smaller than
some constant.
249
TLC, DOE 10104
Add Additional Linear Constraints
.2
.2
.2
.4
.4
.4
.6
.6
.6
.8
.8
.8
X2
X1
X3
Ternary Plot
X
1
<.4
X
2
<.7
X
1
<.5
X
1
>.1
X
3
>0
X
3
<.7
X
2
>.1
X
2
<.7
X
1
<.5
X
1
>.1
X
3
>0
X
3
<.7
X
2
>.1
Open and analyze Extreme Vertices.jmp
250
TLC, DOE 10104
DOE Books:
D.C. Montgomery, Design and Analysis of Experiments. (New York: NY, John Wiley & Sons, 1996).
G.E. Box, J.S. Hunter & W.G. Hunter, Statistics for Experimenters: An Introduction to Design, Data
Analysis, and Model Building (New York: NY, John Wiley & Sons 1978).
Web Sites:
Thomas A. Little Consulting at www.dr-tom.com
Software: SAS JMP contact SAS Institute at www.jmp.com
611 Whitby Lane
Brentwood, CA 94513
1-925-285-1847
drlittle@dr-tom.com
References & Resources
Use of any of the materials contained in these course notes requires the express permission of Thomas A. Little Consulting
Thomas A. Little Consulting, 611 Whitby Lane, Brentwood, CA 94513, September 20, 2003