Design of Experimentsfor Research and Developmentfor Research and Development
Prepared for
Midwest SCCMidwest SCCbyCarr Consulting September 8 2009Carr Consulting September 8, 2009
OutlineOutline
What is DOE?St t f D i d E i t Structure of Designed Experiments
Examples– Factorial Experimentsp– Screening Studies– Optimization Studies– Mixture ExperimentsMixture Experiments
Summary
Carr Consulting 2
What is Design of Experiments?What is Design of Experiments? Design of Experiments (DOE) is: Design of Experiments (DOE) is:
– an efficient, systematic approach» Minimum number of runs to get an answer
R it t k i d» Resource commitment known in advance» Controllable level of precision
– to study the impacts of multiple, controllable factors» Ingredient types and levels» Process conditions» Packaging characteristics
– on key measures of product quality» Objective: Physical/chemical/sensory» Subjective: Consumer acceptance
Carr Consulting 3
» One-at-a-time or simultaneously
When to Use DOE?When to Use DOE?
EARLY -- Screening Studies EARLY -- Screening Studies.– So many variables, so little time.
MIDDLE -- Factorial Experiments.– It’s never that simple.
LATE -- Optimization Studies.Just tell me what’s the best and why?– Just tell me what s the best and why?
Carr Consulting 4
Why Use DOE?Why Use DOE?
Efficiency– Minimum number of samples to get the answer.p g
Sensitivity– “Hidden Replications” add power at no extra cost.
Robust Findings– Effects of each variable are assessed at multiple
levels of all other variables.
Carr Consulting 5
DOE ExamplesDOE Examples
Factorial Experiments Screening Studies Optimization Studies Mixture Experiments
Carr Consulting 6
F t i l E i tFactorial Experiments
ObjectiveObjective
OBJECTIVE: Assess the impact of four production variables on consumers’ acceptance of a shampoovariables on consumers acceptance of a shampoo.– Silicone Type (Type A or Type B)– Silicone Level (0.1% or 2.0%)– Pearlizer (Without or With)– Polymer Level (0.1% vs. 1.0%)
Carr Consulting 8
Technical ApproachTechnical Approach
A factorial experiment comprised of the 16 p ppossible combinations of factor levels was developed to assess the impact of the four formula variablesformula variables.
400 consumers evaluated four of the sixteen samples in a Home Use Test using a BIBD.
Carr Consulting 9
Sixteen Experimental RunsSixteen Experimental Runs
Run Silicone Type Silicone Level Pearlizer Polymer Level OVR Liking1 B 2.0 Without 0.1 812 A 0.1 Without 0.1 583 B 0 1 Without 0 1 503 B 0.1 Without 0.1 504 B 0.1 With 1.0 595 A 2.0 With 1.0 856 A 0.1 With 0.1 587 A 0.1 With 1.0 688 A 2.0 Without 0.1 908 A 2.0 Without 0.1 909 A 0.1 Without 1.0 7110 A 2.0 Without 1.0 9111 B 0.1 With 0.1 5112 B 2.0 With 1.0 7613 B 2.0 Without 1.0 8113 B 2.0 Without 1.0 8114 B 2.0 With 0.1 7815 A 2.0 With 0.1 8516 B 0.1 Without 1.0 62
Carr Consulting 10
Selecting Effects for the ModelSelecting Effects for the Model
One option for fitting the factorial model is to use a One option for fitting the factorial model is to use a probability plot to identify potentially significant effects.
Design-Expert® SoftwareOVR Liking
Shapiro-Wilk testW-value = 0.949p-value = 0.654A: Silicone Type
Half-Normal Plot
ab 95
99
B
Points that fall off the line, high and A: Silicone Type
B: Silicone LevelC: PearlizerD: Polymer Level
Positive Effects Negative Effects
f-N
orm
al %
Pro
ba
70
80
90
95
A
CDBD
line, high and to the right are likely to be significant
Hal
f
0102030
50be significant.
Carr Consulting 11|Standardized Effect|
0.00 5.98 11.96 17.94 23.92
Analysis of Variance Confirms th Si ifi f th Eff tthe Significance of the Effects
Source dfSum of
SquaresMean
Square F-Value P-Valueq qTotal 15 2852.36 A-Silicone Type 1 281.25 281.25 147.5 < 0.0001 B-Silicone Level 1 2289.20 2289.20 1200.3 < 0.0001
C Pearlizer 1 35 70 35 70 18 7 0 0015 C-Pearlizer 1 35.70 35.70 18.7 0.0015 D-Polymer Level 1 106.06 106.06 55.6 < 0.0001 BD 1 121.08 121.08 63.5 < 0.0001Residual 10 19.07 1.91
Carr Consulting 12
Significant Main EffectsSignificant Main Effects
90 90
Ove
rall
Liki
ng
70
80
Ove
rall
Liki
ng
70
80
O
A B
50
60
Without With
50
60
Silicone Type A is more well liked
Without Pearlizer is more well liked (slightly)
Silicone Type Pearlizer
Carr Consulting 13
more well liked. well liked (slightly).
Significant Main EffectsSignificant Main Effects
80
90
80
90
60
70
80
Ove
rall
Liki
ng
60
70
80
Ove
rall
Liki
ng
0.10 0.57 1.05 1.52 2.00
50
60
0.10 0.33 0.55 0.78 1.00
50
60
2.0% Silicone is more well liked than 0 1%
1.0% Polymer is more ll lik d th 0 1%
Silicone Level Polymer Level
Carr Consulting 14
well liked than 0.1%. well liked than 0.1%.
Significant Silicone-by-Polymer InteractionInteraction
2.0% Silicone is more well liked than 0 1% Sili
91.00
0.1% Silicone. At 2.0% silicone,
polymer level does 1.0% Polymer
Liki
ng70 50
80.75
p ynot matter.
At 0.1% Silicone, 1 0% Polymer is
Ove
rall
60.25
70.50
0.1% Polymer1.0% Polymer is more well liked than 0.1% Polymer
0 10 0 57 1 05 1 52 2 00
50.00
Carr Consulting 15
0.10 0.57 1.05 1.52 2.00
Silicone Level
Results & RecommendationsResults & Recommendations
Most well liked shampoo is made at:– Silicone Type A– Silicone Type A– High Silicone Level (2.0 %),– Without Pearlizer, and
Either low or high Polymer (0 1% or 1 0%)– Either low or high Polymer (0.1% or 1.0%). Confirmatory study including both
polymer levels should be conducted.Wh i th dd d t f 1 0%– Why incur the added cost of 1.0% polymer if you do not need to?
Carr Consulting 16
S i E i tScreening Experiments
Screening ExperimentsScreening Experiments
Early stages of researchM f t Many factors
Unknown effects Looking for factors with big effects Looking for factors with big effects Save resources for fine-tuning
experiments run later
Carr Consulting 18
How Screening Studies Save RunsHow Screening Studies Save Runs
F Fi d Si Four-way, Five-way and Six-way interactions are almost always impossible to interpret.Wh i t i
Main Effects and Interactions in a 26 Factorial Experiment
Why invest resources in your experimental designs to be able to estimate effects that you will not be able to understand?
EFFECTNumber of
EffectsNumber of
Runs NeededIntercept 1Main Effects 6 22not be able to understand?
Run a specially selected subset of the full factorial to save resources but still be able to
2-way Interactions 153-way Interactions 204-way Interactions 15 425-way Interactions 66-way Interactions 1resources but still be able to
study the effects that you are interested in.
6 way Interactions 1
Carr Consulting 19
ObjectiveObjective
Assess the impact of six production variables on consumers’ impressions of a sweet snack.p– Diameter (Small vs. Large)– Mouthfeel Ingredient (0.0% vs. 0.4%)
Moisture (2 25% vs 4 50%)– Moisture (2.25% vs. 4.50%)– Particulates (0.0% vs. 0.5%)– Color (3.5% vs. 7.0%)– Mold Position (Open vs. Closed)
Carr Consulting 20
Technical ApproachTechnical Approach
A statistically design variable-screening study comprised of eight experimental sample was p g p pdeveloped to assess the impact of the six production variables
108 consumers evaluated all eight samples 108 consumers evaluated all eight samples. Statistical analyses were conducted to measure
the impact of each product variable on all p pconsumer responses.
Carr Consulting 21
The Experimental DesignThe Experimental Design For each factor, both levels are replicated four times.p Within each level of one factor, the levels of the other
factors are changing. The effect of a factor is not connected to an arbitrary set of initial conditions.connected to an arbitrary set of initial conditions.
Run DiameterMouthfeel Ingredient Moisture Particulates Color
Mold Position
1 S ll 0 4 2 25 0 0 7 0 O1 Small 0.4 2.25 0.0 7.0 Open2 Small 0.0 2.25 0.5 7.0 Closed3 Large 0.4 2.25 0.5 3.5 Open4 Large 0.4 4.50 0.5 7.0 Closed5 Small 0.4 4.50 0.0 3.5 Closed6 Small 0.0 4.50 0.5 3.5 Open7 Large 0.0 2.25 0.0 3.5 Closed8 Large 0.0 4.50 0.0 7.0 Open
Carr Consulting 22
Alternative One at a Time StudyAlternative One-at-a-Time Study Requires one less run but provides only q p y
one comparison for each factor.– Screening Study provides four.
Also, comparisons may be influenced by h i f h b li di ichoice of the baseline conditions.
M thf l M ldRun Diameter
Mouthfeel Ingredient Moisture Particulates Color
Mold Position
1 Small 0.4 2.25 0.0 7.0 Open2 Large 0.4 2.25 0.0 7.0 Open3 Small 0.0 2.25 0.0 7.0 Open4 Small 0.4 4.50 0.0 7.0 Open5 Small 0.4 2.25 0.5 7.0 Open6 Small 0.4 2.25 0.0 3.5 Open7 Small 0.4 2.25 0.0 7.0 Closed
Carr Consulting 23
Analysis is Simple and DirectAnalysis is Simple and DirectColor Liking Size Likingg
babi
lity
95
99
g
abili
ty
95
99
rmal
% P
rob
7080
90
mal
% P
rob
7080
90
95A
F
Hal
f-Nor
0102030
50
Hal
f-Nor
0102030
50
Some responses clearly Other responses clearly|Standardized Effect|
0.00 0.13 0.26 0.39 0.53
|Standardized Effect|0.00 0.48 0.95 1.43 1.90
Carr Consulting 24
Some responses clearly have no significant effects.
Other responses clearly have significant effects.
Results Impact of FactorsResults - Impact of FactorsResponse Diameter
Mouthfeel Ingredient Moisture
Parti-culates Color
Mold Positionp
Acceptance Color Size + - Thickness + - APP Coatingg OVR Appear + - OVR Liking - - - PI - - Flavor
Mouthfeel - - Mouthfeel - - Crispness - - Crunchiness - -Intensity Color + +
Si + Size + - Thick - - Amt Coating + Fruity Flavor Creaminess
Carr Consulting 25
Crispness + + Crunchiness + +
ResultsResults
M ld P i i h d i Mold Position had greatest impact.– Closed is superior to open.
» More well liked overall.» More well liked in key attributes» More well liked in key attributes.
Diameter and Finished Moisture also important.– Larger diameter more well liked for size, thickness and
overall appearanceoverall appearance.– Smaller diameter more well liked overall, for purchase
intent and mouthfeel.– Low finished moisture more well liked for crispness,
hi d llcrunchiness and overall. Mouthfeel Ingredient, Particulates and Color are
not important.
Carr Consulting 26
O ti i ti D iOptimization Designs
Optimization StudiesOptimization Studies
Response Surface Methodology (RSM).– Designed regression analysis.g g y– Built on simple factorial experiments.– Predict responses at points between those
run in the studyrun in the study.» The response surface.
Mixture Experiments– An RSM study in which the levels of all
variables have to sum to a constant.
Carr Consulting 28
RSM ExampleRSM Example
PROJECT OBJECTIVEDetermine the formula for an orange flavored– Determine the formula for an orange flavored beverage with the current flavor system that has the greatest acceptability among consumers.
RESEARCH OBJECTIVE RESEARCH OBJECTIVE– Model the impact of sweetener, acid and flavor
levels on the overall liking for the product in order to d t i th l l th t i ld th t ll lik ddetermine the levels that yield the most well liked product.
Carr Consulting 29
ApproachApproach
Statistically Designed Study– Systematic changes to the levels of sweetener,
acid and flavoracid and flavor.– 15 experimental samples.
Designed Consumer Acceptance Test– Each respondent evaluates 3 of the 15 samples
in a balanced incomplete block design (BIBD).
Analysis y– Link formula and acceptability.– Identify the most well liked formula.
Carr Consulting 30
Three Factor RSMThree-Factor RSM
Carr Consulting 31
Samples and Data
S t A id Fl Liki
Samples and Data
Sweetener Acid Flavor Liking66.2 66.2 66.2 7.266.2 66.2 151 7.066.2 151 66.2 8.566.2 151 66.2 8.566.2 151 151 8.1151 66.2 66.2 9.2151 66.2 151 9.41 1 1 1 66 2 9 8151 151 66.2 9.8151 151 151 10.950 100 100 7.4200 100 100 8.7200 100 100 8.7100 50 100 8.0100 200 100 8.7100 100 50 10.4
Carr Consulting 32
100 100 200 9.9100 100 100 9.3
RSM ModelRSM Model
The Link between Formula and Acceptance
2
2
2
3.02)Swt(LOG63.0
3.02)Swt(LOG78.089.9Liking
2
3.02)Acid(LOG53.0
3.02)Acid(LOG42.0
Response SurfaceResponse Surface
Liking
10 5
8.5
10.5
200%6.5
100%
200%
Acid100%200%
4.5
Carr Consulting 34
50%50%
100%00%
Sweetener
Conto PlotContour Plot
Liking200ci
d
100 10130
Ac 100
88.5
99.5
Sweetener50 100 200
508
150
Carr Consulting 35
Sweetener
ResultsResults
Overall Liking is maximized with a 50% increase Overall Liking is maximized with a 50% increase in sweetener and a 30% increase in acid.
Flavor level has no significant impact on liking.
Carr Consulting 36
SummarySummary
Statistical DOE Statistical DOE– Easy to use.– Applicable at all stages.– Applicable to all research.– Efficient, powerful & rich in information
Carr Consulting 37
Mi t E i tMixture Experiments
Mixture ExperimentMixture Experiment
OBJECTIVE:– Determine the relative proportions of threeDetermine the relative proportions of three
components in a blend that deliver a desired set of sensory properties.
RESEARCH OBJECTIVE:– Model the sensory properties of the product
as a function of its composition. Identify the region of blend ratios within which all action t d d ti fi d i lt lstandards are satisfied simultaneously.
Carr Consulting 39
ApproachApproach
Seven experimental samples:– Three “Pure Blends”. – Three 50:50 Blends.– One 33:33:33 Blend.
Evaluated by 175 assessors using a 7-Pick-4 Evaluated by 175 assessors using a 7 Pick 4 BIBD (yields 100 evaluations/sample).
Sensory responses modeled using proportions of the components as predictorsof the components as predictors.
Carr Consulting 40
Three Component Mixture DesignThree Component Mixture Design
X1
X2 X3
Carr Consulting 41
Samples and DataSamples and Data
X1 X2 X3 Liking100 0 0 5.9
0 100 0 4.70 0 100 4.2
50 50 0 5 850 50 0 5.850 0 50 4.90 50 50 5.6
33 33 33 5 533 33 33 5.5
Carr Consulting 42
Mixture ModelMixture Model
Predictive Model Relating Overall Liking to Blend
Liki 5 9*X1Liking = 5.9*X1+ 4.7*X2+ 4 2*X3+ 4.2 X3+ 1.9*X1*X2- 0.7*X1*X30 3+ 4.5*X2*X3
Carr Consulting 43R2=0.98
Contour Plot of Overall LikingContour Plot of Overall Liking
X1 100DESIGN EXPERT PlotActual Components:
X1=100
Actual Components:X1 = X1X2 = X2X3 = X3
5.75
X2=0X3=0
5.05.5
X2=0X3=0
4.55.0
5.5
X3=100X2=100 X1=0Carr Consulting 44
X3 100X1=0
Optimized Blend RegionOptimized Blend RegionX1=100
DESIGN EXPERT PlotActual Components:X1 = X1
X1 100
X1 = X1X2 = X2X3 = X3
Linger: 3.25 X2=0X3=0Liking: 5.5
g
Flavor Str.: 2.75
Bitterness: 2
X2=0X3 0
O l Pl t
Liking: 5.5
X3=100X1=0X2=100
Carr Consulting 45
Overlay Plot
SummarySummary
Statistical DOE Statistical DOE– Easy to use.– Applicable at all stages.– Applicable to all research.– Efficient, powerful & rich in information
Carr Consulting 46