Vous êtes sur la page 1sur 58

Introduction to Response Surface Methods

1. Response Surface Methodology


Response surface designs
Central composite designs
Whey protein case study

2. Multiple Response Optimization


Whey protein case study

2
Agenda Transition

Screening
Response Surface Methodology: Known
Factors
Unknown
Factors

Response surface designs Screening


Trivial
many

Vital few
Characterization
Central composite designs Factor effects
and interactions

Whey protein case study no


Curvature?

(design and analysis) Optimization


yes
Response
Surface
Methods

Verification no
Confirm? Backup

Celebrate! yes

3
Response Surface Methodology
Subject Matter
Knowledge

Factors
Design of Experiments
Process Region of Operability
Region of Interest
Responses

Empirical Models
(polynomials)

ANOVA

Contour Plots

Optimization
4
Region of Interest
versus Region of Operability

Use factorial design to Region of Interest


get close to the peak.
Then RSM to climb it.

Region of Operability
5
Polynomial Approximations

A decent approximation of any mathematical function can be made via an


infinite series of powers of X, such as that proposed by Taylor. For RSM,
this takes the form:

y 0 1x1 2 x 2 12 x1x 2 11x12 22 x 22

112 x12 x 2 122 x1x 22 111x13 222 x 23 ...

1. The higher the degree of the polynomial, the more closely the Taylor
series can approximate the truth.
2. The smaller the region of interest, the better the approximation. It
often suffices to go only to quadratic level (x to the power of 2).
3. If you need higher than quadratic, think about:
A transformation
Restricting the region of interest
Looking for an outlier(s)
Using a higher order polynomial 6
Simple Maximum (or Minimum)

Maximum
4.00
95

85
2.00
M a x im u m

75

B
0.00
85
65
80
75
70
-2.00
65
4.00 4.00
2.00 2.00
0.00 0.00
-4.00
-2.00 -2.00
B A -4.00 -2.00 0.00 2.00 4.00
-4.00 -4.00

y 83.57 9.39A 7.12B 7.44A 2 3.71B2 5.80AB

7
Rising Ridge

Rising Ridge
4.00
95
90
R is in g R id g e

85
2.00
85

75
80 65
65

B
0.00

65
75

-2.00
70
4.00 4.00
2.00 2.00
0.00 0.00
-4.00
-2.00 -2.00
B A -4.00 -2.00 0.00 2.00 4.00
-4.00 -4.00

y 77.57 8.80A 8.19B 6.95A 2 2.07B2 7.59AB

8
Stationary Ridge

Stationary Ridge
4.00
95
S t a t io n a r y R id g e

85
2.00

75 65
70
80 75
65 75 80 85 85

B
0.00
70
65

-2.00

4.00 4.00
2.00 2.00
0.00 0.00
-4.00
-2.00 -2.00
B A -4.00 -2.00 0.00 2.00 4.00
-4.00 -4.00

y 83.93 10.23A 5.59B 6.95A 2 2.07B2 7.59AB

9
Saddle, or MiniMax

Saddle
4.00
115
155
105
140
95
125
2.00
110
S a d d le

85 65
95 75 75
65 85
80

B
0.00

65
95

105
-2.00
115
4.00 4.00 125
2.00 2.00 135
0.00 0.00 145
-4.00
-2.00 -2.00
B A -4.00 -2.00 0.00 2.00 4.00
-4.00 -4.00

y 84.29 11.06A 4.05B 6.46A 2 0.43B2 9.38AB

10
Response Surface Methodology
Considerations

Requires a quantitative response affected by continuous


factors.
Works best with only a handful of critical factors, those that
survive the screening phases of the experimental program.
Produces an empirical polynomial model which gives an
approximation of the true response surface over a factor
region.
Seeks the optimal settings for process factors so you can
maximize, minimize, or stabilize the responses of interest.
By overlaying contour maps from multiple responses, RSM
can be used to find the ideal "window" of operability.

11
Response Surface Methodology
Types of Designs

Central Composite Design


Classic 5-level design
Great statistical properties

Box Behnken Design


3-level design
Also great statistical properties

Optimal (Custom) Design


Customizable for nearly any situation
Categoric factors, constrained design space

12
Agenda Transition

Screening
Response Surface Methodology: Known
Factors
Unknown
Factors

Response surface designs Screening


Trivial
many

Vital few
Characterization
Central composite designs Factor effects
and interactions

Whey protein case study no


Curvature?

(design and analysis) Optimization


yes
Response
Surface
Methods

Verification no
Confirm? Backup

Celebrate! yes

13
Response Surface Methodology
Central Composite Design

14
Central Composite Design
Elements

Two-level full/fractional factorial (Res V or higher).


Estimate first-order and two factor interactions.

Center points
Estimate pure error and tie blocks together.

Star (or axial) points


Estimate pure quadratic effects.

CCDs are good designs for fitting


second order (quadratic) polynomials
15
Central Composite Design
Template for 3 Factors

A B C
Factorial 1 1 1
points: 1 1 1
1 1 1
1 1 1
1 1 1
1 1 1
1 1 1
1 1 1
Axial (star) 0 0
points: 0 0
0 0
0 0
0 0
0 0
Center 0 0 0
points: 0 0 0
0 0 0
0 0 0
0 0 0
0 0 0 16
Structuring a CCD
Region of Interest

Keep axial Stay within the


(star) runs box* when you
within the use your model
circle. for making
This is the predictions!
region of *region of
operability. interest

17
Agenda Transition

Screening
Response Surface Methodology: Known
Factors
Unknown
Factors

Response surface designs Screening


Trivial
many

Vital few
Characterization
Central composite designs Factor effects
and interactions

Whey protein case study no


Curvature?

(design and analysis) Optimization


yes
Response
Surface
Methods

Verification no
Confirm? Backup

Celebrate! yes

18
Whey Protein Concentrates
Case Study (design and analysis)

Richert et. al.* (1974) used a central composite design to study the effects of five
factors on whey protein concentrates. The factors, with ranges noted in terms of
alpha (star levels), are:
A. Heating temperature, C/30 min. 65 85
B. pH level 4 8
C. Redox potential, volts -0.025 0.375
D. Sodium oxalate, molar 0 0.05
E. Sodium lauryl sulfate, % of solids 0 0.2

The experimenters chose a CCD based on a one-half fraction for the cube portion
(25-1). This rotatable design (with = 2) has six center points.

19
Whey Protein Concentrates
Instructions (1 of 4)

1. The experimenters chose a CCD based on a fraction for the


cube portion (25-1): choose the Fraction.
(Be sure to choose the half fraction before clicking on Enter
factor ranges in terms of alpha.)
2. Then choose Enter factor ranges in terms of alpha.

2
1

Rsm section 3 20
Whey Protein Concentrates
Instructions (2 of 4)

3. The experimenters used a rotatable design ( = 2) and six


center points.

Rsm section 3 21
Whey Protein Concentrates
Instructions (3 of 4)

4. Enter the factor ranges as the alpha values:

5. Enter the one response we will investigate, undenatured


protein, in abbreviated form such as Unde Pro. The units of
measure are percent (%).

Rsm section 3 22
Whey Protein Case Study
Data Factorial portion of CCD

Factor Factor Factor Factor Factor Response


Std Run A:Heat B:pH C:Redox D:Na ox E:Na lau Unde Pro
C / 30 min volt Molar % of soli %
1 9 70.0 5.0 0.075 0.0125 0.15 80.6
2 25 80.0 5.0 0.075 0.0125 0.05 67.9
3 3 70.0 7.0 0.075 0.0125 0.05 83.1
4 19 80.0 7.0 0.075 0.0125 0.15 38.1
5 4 70.0 5.0 0.275 0.0125 0.05 79.7
6 29 80.0 5.0 0.275 0.0125 0.15 74.7
7 22 70.0 7.0 0.275 0.0125 0.15 71.2
8 18 80.0 7.0 0.275 0.0125 0.05 36.8
9 11 70.0 5.0 0.075 0.0375 0.05 81.7
10 31 80.0 5.0 0.075 0.0375 0.15 66.8
11 2 70.0 7.0 0.075 0.0375 0.15 73.0
12 23 80.0 7.0 0.075 0.0375 0.05 40.5
13 13 70.0 5.0 0.275 0.0375 0.15 74.9
14 30 80.0 5.0 0.275 0.0375 0.05 74.2
15 7 70.0 7.0 0.275 0.0375 0.05 63.5
16 12 80.0 7.0 0.275 0.0375 0.15 42.8

23
Whey Protein Case Study
Data Star and center points

Factor Factor Factor Factor Factor Response


Std Run A:Heat B:pH C:Redox D:Na ox E:Na lau Unde Pro
C / 30 min volt Molar % of soli %
17 8 65.0 6.0 0.175 0.0250 0.10 80.9
18 27 85.0 6.0 0.175 0.0250 0.10 42.4
19 16 75.0 4.0 0.175 0.0250 0.10 73.4
20 24 75.0 8.0 0.175 0.0250 0.10 45.0
21 10 75.0 6.0 -0.025 0.0250 0.10 66.0
22 17 75.0 6.0 0.375 0.0250 0.10 71.7
23 15 75.0 6.0 0.175 0.0000 0.10 77.5
24 28 75.0 6.0 0.175 0.0500 0.10 76.3
25 32 75.0 6.0 0.175 0.0250 0.00 67.4
26 21 75.0 6.0 0.175 0.0250 0.20 86.5
27 20 75.0 6.0 0.175 0.0250 0.10 77.4
28 5 75.0 6.0 0.175 0.0250 0.10 74.6
29 6 75.0 6.0 0.175 0.0250 0.10 79.8
30 26 75.0 6.0 0.175 0.0250 0.10 78.3
31 1 75.0 6.0 0.175 0.0250 0.10 74.8
32 14 75.0 6.0 0.175 0.0250 0.10 80.9

24
Case Study
Whey Protein Concentrates

There were nine responses, lets look at three key ones:

Y1 Undenatured protein, %.

Y2 Whipping time, min.

Y3 Time at first drop, min.

25
Whey Protein Case Study
Sequential Model Sum of Squares

Sequential Model Sum of Squares


Sum of Mean F
Source Squares DF Square Value Prob > F
Mean 1.516E+005 1 1.516E+005
Linear 4323.77 5 864.75 9.77 < 0.0001
2FI 883.30 10 88.33 1.00 0.4848
Quadratic 1179.84 5 235.97 10.88 0.0006 Suggested
Cubic 202.04 5 40.41 6.64 0.0196 Aliased
Residual 36.51 6 6.09
Total 1.582E+005 32 4943.93

"Sequential Model Sum of Squares": Select the highest order polynomial where the
additional terms are significant.

26
Whey Protein Case Study
Lack of Fit Tests

Lack of Fit Tests


Sum of Mean F
Source Squares DF Square Value Prob > F
Linear 2268.60 21 108.03 16.32 0.0029
2FI 1385.30 11 125.94 19.03 0.0022
Quadratic 205.46 6 34.24 5.17 0.0459 Suggested
Cubic 3.42 1 3.42 0.52 0.5044 Aliased
Pure Error 33.09 5 6.62

"Lack of Fit Tests": Want the selected model to have insignificant lack-of-fit.

Do you want significant lack of fit?

27
Whey Protein Case Study
Model Summary Statistics

Model Summary Statistics


Std. Adjusted Predicted
Source Dev. R-Squared R-Squared R-Squared PRESS
Linear 9.41 0.6526 0.5858 0.4673 3529.61
2FI 9.42 0.7859 0.5852 -1.1703 14379.16
Quadratic 4.66 0.9640 0.8985 0.1632 5544.20 Suggested
Cubic 2.47 0.9945 0.9715 0.4325 3759.87 Aliased

"Model Summary Statistics": Focus on the model minimizing the "PRESS", or


equivalently maximizing the "PRED R-SQR".

Whats wrong with these statistics?

28
Whey Protein Case Study
Significance (?) of Quadratic Terms

Lets try reducing this model to only significant terms.


Source SS DF MS F Prob > F
A 2458.35 1 2458.35 113.36 < 0.0001
B 1807.87 1 1807.87 83.36 < 0.0001
C 0.26 1 0.26 0.012 0.9147
D 12.18 1 12.18 0.56 0.4693
E 45.10 1 45.10 2.08 0.1771
A2 506.85 1 506.85 23.37 0.0005
B2 667.23 1 667.23 30.77 0.0002
C 2 162.93 1 162.93 7.51 0.0192
D 2 3.48 1 3.48 0.16 0.6965
E2 3.23 1 3.23 0.15 0.7069
AB 616.28 1 616.28 28.42 0.0002
AC 122.66 1 122.66 5.66 0.0366
AD 50.06 1 50.06 2.31 0.1569
AE 7.98 1 7.98 0.37 0.5564
BC 45.23 1 45.23 2.09 0.1766
BD 1.05 1 1.05 0.048 0.8298
BE 3.71 1 3.71 0.17 0.6873
CD 0.031 1 0.031 1.412E-003 0.9707
CE 36.30 1 36.30 1.67 0.2222
DE 0.016 1 0.016 7.205E-004 0.9791 29
Algorithmic Model Reduction
Backward Selection

1. Begin with the full model.


2. Remove from the model the factor with the smallest F value.
3. Stop when the p-value of the next factor out satisfies the
specified alpha value criterion.

We put this first on the list because it gives every term a


chance to get into the model.

30
Whey Protein Case Study
Model Reduction (Instructor-led)

1. Return to model selection by


pressing the Model button
2. Reduce the model by changing the
Selection method from Manual to
Backward.
3. Choose ANOVA for this reduced
model.

31
Hierarchical Models*

Y E S!

32
Whey Protein Case Study
Full vs Reduced Quadratic Model (1 of 2)

ANOVA for Response Surface Quadratic Model (Full)


Sum of Mean F
Source Squares DF Square Value Prob > F
Model 6386.91 20 319.35 14.73 < 0.0001
Residual 238.55 11 21.69
Lack of Fit 205.46 6 34.24 5.17 0.0459
Pure Error 33.09 5 6.62
Cor. Total 6625.46 31

ANOVA for Response Surface Reduced Quadratic Model


Sum of Mean F
Source Squares DF Square Value Prob > F
Model 6331.63 12 527.64 34.12 < 0.0001
Residual 293.83 19 15.46
Lack of Fit 260.73 14 18.62 2.81 0.1297
Pure Error 33.09 5 6.62
Cor. Total 6625.46 31

33
Whey Protein Case Study
Full vs Reduced Quadratic Model (2 of 2)

Full Quadratic Model


Std. Dev. 4.66 R-Squared 0.9640
Dep Mean 68.83 Adj R-Squared 0.8985
C.V. 6.77 Pred R-Squared 0.1632
PRESS 5544.20 Adeq. Precision 11.789

Reduced Quadratic Model


Std. Dev. 3.93 R-Squared 0.9557
Mean 68.83 Adj R-Squared 0.9276
C.V. % 5.71 Pred R-Squared 0.8589
PRESS 934.73 Adeq Precision 17.589

Benefits are clear for using the reduced model for this response.

34
Case Study
Whey Protein Concentrates

There were nine responses, lets look at three key ones:

Y1 Undenatured protein, %.

Y2 Whipping time, min.

Y3 Time at first drop, min.

35
Whey Protein Concentrates
Y2 Whipping time

Response: Whip time Transform: Base 10 log Constant: 0.000


ANOVA for Response Surface Reduced Quadratic Model
Analysis of variance table [Partial sum of squares]
Sum of Mean F
Source Squares DF Square Value Prob > F
Model 0.47 7 0.067 15.43 < 0.0001
A 0.23 1 0.23 53.78 < 0.0001
B 5.528E-004 1 5.528E-004 0.13 0.7241
C 0.037 1 0.037 8.53 0.0075
D 4.136E-003 1 4.136E-003 0.95 0.3384
A2 0.100 1 0.100 23.03 < 0.0001
AB 0.069 1 0.069 16.02 0.0005
BD 0.024 1 0.024 5.56 0.0269
Residual 0.10 24 4.333E-003
Lack of Fit 0.093 19 4.905E-003 2.27 0.1854
Pure Error 0.011 5 2.161E-003
Cor Total 0.57 31

36
Whey Protein Concentrates
Y2 Whipping time

Response: Whip time Transform: Base 10 log Constant: 0.000

Std. Dev. 0.066 R-Squared 0.8182


Mean 0.64 Adj R-Squared 0.7652
C.V. 10.28 Pred R-Squared 0.6544
PRESS 0.20 Adeq Precision 18.865

37
Case Study
Whey Protein Concentrates

There were nine responses, lets look at three key ones:

Y1 Undenatured protein, %.

Y2 Whipping time, min.

Y3 Time at first drop, min.

38
Whey Protein Concentrates
Y3 Time at first drop

Response: Time at first drop Transform: Base 10 log Constant: 0.000


ANOVA for Response Surface Reduced 2FI Model
Analysis of variance table [Partial sum of squares]
Sum of Mean F
Source Squares DFSquare Value Prob > F
Model 0.85 7 0.12 8.01 < 0.0001
A 0.019 1 0.019 1.26 0.2735
B 0.53 1 0.53 35.41 < 0.0001
C 0.089 1 0.089 5.89 0.0231
D 2.095E-004 1 2.095E-004 0.014 0.9072
E 0.025 1 0.025 1.68 0.2075
AD 0.11 1 0.11 7.55 0.0112
AE 0.064 1 0.064 4.24 0.0504
Residual 0.36 24 0.015
Lack of Fit 0.35 19 0.019 11.72 0.0063
Pure Error 7.950E-003 5 1.590E-003
Cor Total 1.21 31

39
Whey Protein Concentrates
Y3 Time at first drop

Response: Time at first drop Transform: Base 10 log Constant: 0.000

Std. Dev. 0.12 R-Squared 0.7001


Mean 1.00 Adj R-Squared 0.6127
C.V. 12.34 Pred R-Squared 0.3933
PRESS 0.73 Adeq Precision 12.609

40
Whey Protein Concentrates
Optimization

Next Step:
Use the three response models we just fit to find the best
tradeoff in properties to give the optimum operating
conditions.

41
Introduction to Design of Experiments

1. Response Surface Methodology


Response surface designs
Central composite designs
Whey protein case study

2. Multiple Response Optimization


Whey protein case study

42
Agenda Transition

Multiple Response Optimization:


Screening
Whey protein case study Known
Factors
Unknown
Factors

(optimization) Trivial
Screening many

Vital few
Characterization
Factor effects
and interactions

no
Curvature?

yes
Optimization Response
Surface
Methods

Verification no
Confirm? Backup

Celebrate! yes

43
Simultaneous Optimization
of Multiple Responses

1. Analyze each response separately and establish an appropriate


transformation and model for each.
2. Optimize using the models to search the independent factor
space for a region that simultaneously satisfies the
requirements placed on the responses.

Useful models are essential!


Design of experiments is critical!

44
First Step: Develop Good Models
Dont Over Interpret the Statistics!

Be sure the fitted surface adequately represents your process


before you use it for optimization. Check for:
1. A significant model: Large F-value with p<0.05.
2. Insignificant lack-of-fit: F-value with p>0.10.
3. Adequate precision >4.
4. Well behaved residuals: Check diagnostic plots!

45
Whey Protein Concentrates
Y1 Undenatured protein

Response: Undenatured Protein


ANOVA for Response Surface Reduced Quadratic Model
Sum of Mean F
Source Squares DF Square Value Prob > F
Model 6331.63 12 527.64 34.12 < 0.0001
Residual 293.83 19 15.46
Lack of Fit 260.73 14 18.62 2.81 0.1297
Pure Error 33.09 5 6.62
Cor. Total 6625.46 31

Std. Dev. 3.93 R-Squared 0.9557


Mean 68.83 Adj R-Squared 0.9276
C.V. % 5.71 Pred R-Squared 0.8589
PRESS 934.73 Adeq Precision 17.589

46
Whey Protein Concentrates
Y2 Whipping time

Response: Whip time Transform: Base 10 log Constant: 0.000


ANOVA for Response Surface Reduced Quadratic Model
Sum of Mean F
Source Squares DF Square Value Prob > F
Model 0.47 7 0.067 15.43 < 0.0001
Residual 0.10 24 4.333E-003
Lack of Fit 0.093 19 4.905E-003 2.27 0.1854
Pure Error 0.011 5 2.161E-003
Cor Total 0.57 31

Std. Dev. 0.066 R-Squared 0.8182


Mean 0.64 Adj R-Squared 0.7652
C.V. 10.28 Pred R-Squared 0.6544
PRESS 0.20 Adeq Precision 18.865

47
Whey Protein Concentrates
Y3 Time at first drop

Response: Time at first drop Transform: Base 10 log Constant: 0.000


ANOVA for Response Surface Reduced Quadratic Model
Sum of Mean F
Source Squares DF Square Value Prob > F
Model 0.85 7 0.12 8.01 < 0.0001
Residual 0.36 24 0.015
Lack of Fit 0.35 19 0.019 11.72 0.0063
Pure Error 0.00795 5 1.590E-003
Cor Total 1.21 31

Std. Dev. 0.12 R-Squared 0.7001


Mean 1.00 Adj R-Squared 0.6127
C.V. 12.34 Pred R-Squared 0.3933
PRESS 0.73 Adeq Precision 12.609

48
Response Surface Numeric Optimization
Desirability as an Objective Function (1/2)

To determine a best combination of responses, we use an


objective function, D(X), that involves the use of a geometric
mean:
1
1
n

D d1 d2 ... dn
n
n di
i 1
The di, which range from 0 to 1 (least to most desirable
respectively), represents the desirability of each individual
(i) response.
n is the number of responses being optimized.

49
Response Surface Numeric Optimization
Desirability as an Objective Function (2/2)

Now you can search for the greatest overall desirability (D) for
responses and/or factors (for example, if time is a factor, you may
want to keep it to a minimum):
D = 1 indicates that all the goals are satisfied.
(If this happens, youre probably not asking
for enough!)
D = 0 when one or more responses fall outside acceptable
limits. (Hopefully this will not happen, but if so, try relaxing
some of your criteria!)

50
Desirability as an Objective Function
Assigning Optimization Parameters (1/2)

The crucial phase of numerical optimization is assignment of


various parameters that define the application of individual
desirabilities (dis). The most important are:
Goal (none, maximum, minimum, target or range)
Limits (lower and upper).
In this case:
Want to maximize undenatured protein.
Want to minimize whip time.
Want to maximize time at first drop.

51
Desirability as an Objective Function
Assigning Optimization Parameters (2/2)

Of lesser importance are the parameters:


Weight (0.1 to 10) (Well leave them all = 1)
Importance (5-point scale displayed + to +++++)

In this case:
Undenatured protein is most important, + + + + +.
Whip time is least important, + +.
Time at first drop, this is of intermediate importance, + + +.

52
Whey Protein Concentrates
Optimization

Want to maximize
undenatured protein,
this is the most important
response:
+++++

53
Whey Protein Concentrates
Optimization

Want to minimize whip


time, this is the least
important response:
++

54
Whey Protein Concentrates
Optimization

Want to maximize time at


first drop, this is of
intermediate importance:
+++

55
Whey Protein Concentrates
Numeric Optimization

Solutions
# A B C D E Y1 Y2 Y3 D
1 70.00 6.23 0.15 0.04 0.15 82.948 3.3895 12.7 0.217
2 70.00 6.24 0.15 0.04 0.15 82.917 3.3842 12.7 0.217
3 70.00 6.26 0.14 0.04 0.15 82.924 3.3902 12.6 0.214
4 70.00 6.15 0.17 0.04 0.15 82.852 3.3890 12.7 0.214
5 70.00 6.17 0.16 0.04 0.14 82.921 3.3925 12.5 0.212

Factor Name Response Name


A Heating Y1 Undenatured Protein
B pH Y2 Whip time
C Redox pot Y3 Time at first drop
D Na oxalate
E Na lauryl

56
Whey Protein Concentrates
Numeric Optimization

57
Summary Response Surface Methods

Goal Optimization of process


Tools
Central Composite design
(when it fits the problem)
Optimal (Custom) design if needed
(Watch for webinar - Part 2!)
Numerical Optimization

58
Practical Paperbacks on DOE*
by Mark Anderson and Pat Whitcomb

User Review of DOE Simplified:

As an engineer (just beginning self study on the topic of DOE) I found this book
very useful. The authors provide practical insight that I was unable to find in other
DOE or statistics books. This is not a book for advanced statisticians, however, it is
a great book for someone trying to understand and apply the principles of DOE.
* Published by Productivity Press, New York. 59

Vous aimerez peut-être aussi