...

Optimize Your Process-Optimization Efforts

by user

on
Category: Documents
5

views

Report

Comments

Transcript

Optimize Your Process-Optimization Efforts
1
Optimize Your Process-Optimization Efforts
Highly efficient, statistically based methods can identify the vital few
factors that affect process efficiency and product quality.
Mark J. Anderson and Patrick J. Whitcomb
Stat-Ease, Inc.
2021 East Hennepin Ave., Suite 480, Minneapolis, MN 55413
Telephone: 612.378.9449, Fax: 612.378.2152
(Writer’s note: this is the first of a two-part series on design of experiments. Part
two provides details on a relatively new technique called mixture design -- a powerful
technique for improving formulations.)
What would you do if confronted with “opportunities” like these?
1. A chemist synthesizes a new wonder compound in the laboratory. The
marketing people anticipate big profits but only if the product can be made economically
at high yields. The chemist provides a very basic recipe sheet with suggested conditions.
Your job is to scale up the process as quickly as possible.
2. A competitor makes a minor, but noticeable, improvement to their product
while simultaneously reducing price. The sales group fears that your product may get
knocked out of the market. Your job is to fine tune the existing plant process and get
more yield with better product quality.
In either case, you probably would first try to gather fellow experts and identify all
possible variables that may affect yields and product quality. An exhaustive list might
include dozens of potential factors - many more than you could possibly investigate.
Some of these variables can’t be controlled: Be sure to record their values. Other
variables won’t be given much priority: Hold these at fixed levels. Still, you’re likely to
be left with five or more possible control factors. Now what do you do?
The traditional approach to experimentation requires you to change only one
factor at a time (OFAT). However, the OFAT approach doesn’t provide data on
interactions of factors, a likely occurrence with chemical processes. An alternative
approach called “two-level factorial design” can uncover critical interactions. This
statistically based method involves simultaneous adjustment of experimental factors at
only two levels: high and low. The two-level design offers a parallel testing scheme
that’s much more efficient than the serial approach of OFAT.
By restricting the tests to only two levels, you minimize the number of
experiments. The contrast between levels gives you the necessary driving force for
process improvement. You don’t need to run three levels of every factor until you get
close to the optimum. At the beginning of your investigation many factors must be
considered. Doing all combinations of every factor at three levels would produce a
prohibitively large number of runs.
2
Strategy of experimentation
The basics of two-level factorial design are well-documented.1,2 To illustrate their
application to chemical engineering, we will investigate a process optimization of
waterborne polyurethane.3 The chemical engineers who performed the experiments
wanted small particles to ensure a stable aqueous dispersion. This case study illustrates a
very basic two-phase strategy for experimenters:
Phase 1. Use two-level factorial designs as screening tools to separate the vital
few factors (including interactions) from the trivial many that have no
significant impact.
Phase 2. Follow up by doing an in-depth investigation of the surviving factors.
Generate a “response surface” map4 and move the process to the
optimum location.
Part two of this series of articles on design of experiments will cover the
optimization via response surface methods (RSM). This article focuses on the screening
phase. It presents the vital tools of statistical design of experiments (DOE) via two-level
factorial design. We will do this from an engineering perspective, with an emphasis on
the practical aspects.
Fractional factorials for maximum efficiency
When screening factors, you need not run the full combination of high and low
levels: often a fraction will do. The polyurethane experimenters chose to run a half
fraction. They studied five factors (Table 1) in sixteen experiments which can be
symbolized mathematically as 25-1. It provides sound estimates of main effects and twofactor interactions if you can assume that three-factor or higher interactions will not
occur. Generally this is a safe assumption, but you should always make confirmation
runs to verify experimental findings. Table 2 shows high-resolution design options for
five or more factors. (Run full factorials for four or fewer factors if you want high
resolution of effects). The designs can be constructed with the aid of a textbook, or better
yet, with a statistical software package, most of which offer design of experiments
capabilities5,6.
You will find designs available with as little as k+1 runs, where k equals the
number of factors you want to test. For example you could test 7 factors in 8 runs, or 15
factors in 16 runs. However, these “saturated” designs provide very poor resolution:
main effects will be confused with two-factor interactions. We advise that you avoid
running such low resolution designs.
The specific design layout for the polyurethane (PU) case is shown in Table 3.
Columns A through D are laid out according to a standard order that can be obtained from
any textbook on design of experiments. This portion of the matrix represents a full twolevel design for four factors in sixteen runs. In order to get the additional factor E into the
design, the four-factor interaction ABCD is used. Multiply the A, B, C and D columns
and you find the product in column E:
E = ABCD
(1)
Statisticians call this relation between E and ABCD an “alias”. You cannot
differentiate between the two: an observed difference in response due to a change in E
could really be caused by ABCD, or even some combination of E and ABCD. But as
3
discussed earlier, you can ignore three factor or higher interactions, so don’t worry. It
turns out that with the half fraction, every effect will have another effect aliased with it.
These are shown under the column labeled “Alias” in Table 4. All of these aliased effects
are third order or higher so they can be ignored. If you run a 1/4th fraction, each effect
will have three aliases, effects in an 1/8th fraction each have seven aliases, and so on.
Note the balanced array of plus (high) and minus (low) levels in the test matrix
(Table 3). Each column contains eight pluses and eight minuses. The matrix offers a
very important statistical property called “orthogonality” which means that factors are not
correlated. If you just collected happenstance data from a plant, it is highly unlikely you
would get an array of factors like this. You would probably find that factors such as
temperature and pressure go up and down together. As factors become more and more
correlated, the error in estimation of their effects becomes larger and larger. That’s not
good.
Orthogonal test matrices make effect estimation neat and easy. For example, the
effect of factor D is calculated by simply averaging the responses at the plus level and
subtracting the average at the minus levels.
Effect = Mean Y+ - Mean Y−
(2)
Applying a transformation to satisfy statistical assumptions
Notice in Table 3 that the response varies by nearly an order of magnitude, from
40 to 389 nanometers. In situations like this, statisticians routinely perform a
transformation of the response, most commonly with a logarithm. Chemical engineering
students do the same thing when they use special graph paper, such as log scale, to get the
data to come out in a straight line. Contractive transformations, such as the log or square
root, counteract a very common relationship: the true standard deviation (σ) increases as
the true mean (µ) goes up. Statisticians express this as a power law1:
σ = fn(µα)
(3)
In the ideal case, there will be no relationship between standard deviation and the
mean, so the power (α) will be zero. This satisfies an important statistical assumption constant variance. If you cannot satisfy this assumption, consider a transformation.
Ideally, your engineering or chemistry knowledge will guide you in the selection of an
appropriate transformation. However, if you can’t predict what the relationship should
be, just try the log or square root or some other transformation. We got a better statistical
fit from the square root transformation. The resulting responses can be seen in the last
column of Table 4. Later in this article we will discuss how to validate your model,
whether transformed or not, via analysis of the residuals.
Using statistical principles to pick significant factors
You now know how to calculate effects. It seems obvious that you should be pick
the largest ones and run with those - right? Wrong! How do you know where to make
the cut off? What if none of the effects are real, and you’ve just measured results due to
random error? Somehow the vital few significant factors must be screened out of the
trivial many that occur due to chance. You can do this easily with a graph called a “halfnormal plot”. Simply rank the absolute value of the effects from low to high. Then
assign cumulative probability values according to the following formula.
4
i − 0.5
)
(4)
m
where i = rank
m = number of effects
Now plot the effects versus their assigned probability on half-normal graph paper.
Then find the group of near-zero effects and draw a line through them, but don’t include
the points after the bend in the ‘dogleg’. Anything significant will fall off to the right of
the line. Figure 1 shows the half-normal plot of effects for the polyurethane case.
Significant effects are labeled. The near-zero effects fall on a straight line - exhibiting
normal scatter. These insignificant effects can be used to estimate experimental error.
If you want to be conservative, consider replicating the design to get estimates of
“pure” error. Be sure that you go back through all of the steps, e.g., charge the reactor,
bring it up to temperature, line it out, take samples and do the analysis. Don’t just reanalyze or re-sample and put these in as replicated because you won’t get a fair estimate
of the pure error. Also, be sure to randomize the run order of your entire design,
including replicates. Otherwise you leave yourself open to “lurking factors”, such as
ambient temperature or catalyst degradation, that could confound your factor estimates.
Another way to add replication is to put a “centerpoint” in your design. This is a
set of conditions at the midpoint of every factor level. For example the centerpoint for
the polyurethane experiment would be at 75 ppm catalyst, 3.65 acetone/PU, 40 degrees
Centigrade, 925 rpm and 3 milliliters per minute addition rate. Most experimenters
repeat the centerpoint several times mixed in randomly with the remaining design points.
In addition to an estimate of pure error from the replication, you can then estimate
“curvature” in your system. Significant curvature indicates that your response behaves in
a non-linear fashion. Then you will need to run additional factor levels and employ
response surface methods. Issues related to curvature will be discussed in more detail in
part two of this series on DOE. In most cases you will find that curvature is not
significant, which means that you can rely on the two-level design.
Given a valid estimate of experimental error, regardless of the source, standard
statistical analyses can then be performed to validate the overall outcome and individual
effects. Textbooks provide hand-calculation schemes for doing statistical analysis of
two-level factorials, but it’s much easier to let a statistical software program to do this
work for you.
Be sure you choose a program that provides “residual analysis” capabilities.
Residuals are the difference between actual and predicted response. The residuals must
be approximately normal. You can check this by plotting residuals on normal or halfnormal paper. Figure 2 shows the normal plot of residuals for the polyurethane case. The
plot was constructed by ranking the residuals from low to high and assigning probabilities
according to Equation 4. A crude but effective way to evaluate this plot is the “pencil
test”: If you can cover the points with your pencil, then the residuals are well-behaved. In
this case, they look fine.
Your statistical software also should offer a plot of residuals versus predicted
level such as that shown in Figure 3. This plot exhibits a desirable scatter - no obvious
change in variation as the level increases.
Pi = 100 (
5
If you do see bad patterns on either of these residual plots, such as an “S” shape
on the normal plot (Figure 2) or a megaphone shape on the residual versus predicted plot
(Figure 3), consider the use of a response transformation. The log transformation often
helps. You might also try a square root or one of many other functions. Refer to a DOE
textbook for statistical advice on this subject. However, there is no substitute for your
process knowledge. This should guide you in selection of a transformation.
Residual analysis also may reveal individual outliers. But be careful, don’t delete
points unless you can assign a special cause, such as a temporary breakdown in an
agitator or the like. Quite often a outlier turns out to be simply an error in data entry.
Digits can get transposed very easily.
Interpreting the results
Now you are ready to make your report. Start by making a plot of any significant
main effects that are not part of a significant interaction. In the polyurethane case, only
factor E stands alone: It does not combine with any other factors in a significant
interaction. The effect of factor E can be seen in Figure 4. Remember that each point in
this plot represents a contrast between the response averages at low versus high factor
levels. In this case there are eight runs at each level, so the results carry a lot of weight.
Clearly factor E needs to be set at its low level to get minimal particle size. Next,
produce the interaction plots. In this case we found three interactions active: AC, BC and
BD, which can be seen in Figures 5,6 and 7, respectively. Notice that the lines on these
plots are not parallel. In other words, the effect of one factor depends on the level of the
other, so it would be inappropriate to display any of these factors by themselves. For
example, on the interaction plot in Figure 7, notice that factor B (acetone/PU ratio) has a
much bigger impact when D (agitation) is at the higher rate. Clearly it’s best to go with
low B and high D. Then from the BC plot (Figure 6) it’s clear that factor C should be set
at its high level. Finally, from the AC plot (Figure 5), given that C will be set at its plus
level, it makes no difference how you set A, just pick a convenient level. Table 5
summarizes the recommended settings for minimizing particle size.
Before you make a final recommendation on the new factor levels, it would be
wise to perform confirmation runs. You can predict the outcome with a simple equation
that uses the overall average modified up or down depending on the level you set each
factor to. Statisticians call this a “coded” equation because you plug in values of plus
one for high and minus one for low levels. (A midpoint setting is entered as zero.)
p-1
Effect i
Predicted Response = Overall Average + ∑ (
)X i
(5)
2
i=1
For the polyurethane case the predictive model is:
Y1/2 = 11.15 - 0.76A + 2.04B - 1.91C + 0.06D + 1.39E
+ 0.74AC - 0.95BC + 1.02BD
(6)
This model includes factor D to maintain hierarchy of terms: D is one of the
“parent” terms for the significant BD interaction. Although factor D may not be
important alone, you can see from the BD interaction plot that it does make a difference depending on the level of B. When B is low, increasing D causes smaller particle size.
Conversely, when B is high, increasing D causes larger particle size. These “crossover”
type interactions can cause confusion for OFAT experimenters. By using two-level
6
factorial designs you can screen out these nuggets of information and perhaps accomplish
a real breakthrough.
Plugging in the recommended settings in coded form gives a predicted outcome.
(Let’s choose the high level for A).
Y1/2 = 11.15 - 0.76(+1) + 2.04(-1) - 1.91(+1) + 0.06(+1) + 1.39(-1)
+ 0.74(+1)(+1) - 0.95(-1)(+1) + 1.02(-1)(+1)
= 5.78
(7)
Then to get the response back to the original units of measure, the transformation
must be reversed.
Predicted PU particle size = 5.783= 33 nanometers
This compares well with the observed result of 40 for experiment number 14,
which happened to be run at the recommended settings.
What’s in it for you
The case study on polyurethane illustrates how two-level factorials can be applied
to a chemical process with many variables. The design of experiments uncovered several
interactions which led to a breakthrough product improvement. The experimenters found
that one of the factors, catalyst concentration (A), could be eliminated from further
consideration. In part two of this series on DOE, we will take a look at an in-depth
optimization on three of the factors critical to the waterborne polyurethane system. The
article shows how to set up and analyze a response surface method (RSM) design. Figure
8 shows an example of a three-dimensional response surface graph with contours
projected underneath.
If you equip yourself with the basic tools of statistical DOE, you will be in a
position to make the most of opportunities such as those presented at the outset of this
article. Your reputation will be enhanced and the competitive position of your company
advanced.
Nomenclature
Number of factors
Number of model parameters, including intercept
Polyurethane
Factor
Response
k
=
p
=
PU
=
X
=
Y
=
Greek Letters
α
=
Exponent for power law
µ
=
True mean
σ
=
True standard deviation
7
Literature Cited
(1)
(2)
(3)
(4)
(5)
(6)
Box, G.E.P., Hunter, W.G., and Hunter, J.S., Statistics for Experimenters, John Wiley & Sons, Inc,
New York, 1978.
Montgomery, D.C., Design and Analysis of Experiments, 3rd ed., John Wiley & Sons, Inc, New
York, 1991.
Yang, C.H., Lin, S.M. and Wen, T.C., “Application of Statistical Experimental Strategies to the
Process Optimization of Waterborne Urethane,” Polymer Engineering and Science, Vol. 35, No. 8
(April 1995).
Myers, R.H., Montgomery, D.C., Response Surface Methodology, John Wiley & Sons, Inc, New
York, 1995.
“Mathematics, Statistics” section, 1995 CEP Software Directory, pages 27-30 (supplement to
December 1994 issue of Chemical Engineering Progress).
Helseth, T.J., et al, Design-Ease, Version 3 for Windows or Macintosh, Stat-Ease, Inc,
Minneapolis, 1994 ($395).
8
Table 1. Factors and Levels for Polyurethane (PU) Experiment
Factor
Description
Low Level
High Level
−
A (X1)
B (X2)
C (X3)
D (X4)
E (X5)
Catalyst (ppm)
Acetone/PU (ratio)
Phase-inversion temperature (ºC)
Agitation rate (rpm)
Water-addition rate (mL/min)
0
2.8
30
350
2.0
+
150
4.5
50
1500
4.0
Table 2. High Resolution Two-Level Fractional Factorials*
Factors
# Runs
Fraction
# Runs
Full
Applied
2 Level
Fractional
Design
Design
5
32 (25)
1/2 (2-1)
16 (25-1)
6
64 (26)
1/2 (2-1)
32 (26-1)
7
128 (27)
1/4 (2-2)
32 (27-2)
8
256 (28)
1/8 (2-3)
32 (28-3)
9
512 (29)
1/8 (2-3)
64 (28-3)
10
1024 (210)
1/16 (2-4)
64 (210-4)
11
2048 (211)
1/32 (2-5)
64 (211-5)
*These designs give clean estimates of all main effects and all, or nearly all, twofactor interactions. For details on construction and properties of these designs see
Box, Hunter and Hunter1 Chapter 12.
9
Table 3. 25-1 Design Matrix and Data for Polyurethane Case
Standard
A
B
C
D
E
Particle Size
Order
(=ABCD)
(nanometers)
Square root
Particle Size
1
−1
−1
−1
−1
+1
196
14.00
2
+1
−1
−1
−1
−1
70
8.37
3
−1
+1
−1
−1
−1
234
15.30
4
+1
+1
−1
−1
+1
257
16.03
5
−1
−1
+1
−1
−1
75
8.66
6
+1
−1
+1
−1
+1
86
9.27
7
−1
+1
+1
−1
+1
101
10.04
8
+1
+1
+1
−1
−1
50
7.07
9
−1
−1
−1
+1
−1
85
9.21
10
+1
−1
−1
+1
+1
76
8.72
11
−1
+1
−1
+1
+1
389
19.72
12
+1
+1
−1
+1
−1
173
13.15
13
−1
−1
+1
+1
+1
70
8.36
14
+1
−1
+1
+1
−1
40
6.32
15
−1
+1
+1
+1
−1
100
10.00
16
+1
+1
+1
+1
+1
202
14.21
10
Table 4. Estimates of Effects for Polyurethane Case
Alias
Effect
(Square root scale)
A:Catalyst conc*
BCDE
-1.52
B:Acetone/PU*
ACDE
4.08
C:PI Temp*
ABDE
-3.82
D:Agitation
ABCE
0.12
E:Water add*
ABCD
2.79
AB
CDE
0.37
AC*
BDE
1.47
AD
BCE
0.29
AE
BCD
0.54
BC*
ADE
-1.90
BD*
ACE
2.04
BE
ACD
0.84
CD
ABE
0.84
CE
ABD
-0.32
DE
ABC
0.29
*Significant effects based on inspection of normal plot. Statistical analysis
revealed that each of these were significant at the 5% level (>95% confidence).
Table 5. Recommended Settings for Minimal Particle Size
Factor
Best Setting
Coded Level
A:Catalyst conc
Pick a level*
−1 to +1
B:Acetone/PU
Low
−1
C:PI Temp
High
+1
D:Agitation
High
+1
E:Water add
Low
−1
*(Factor A not statistically significant factor when factor C set at high level)
11
99
Half Normal % probability
97
B
95
90
C
85
E
80
BD
70
BC
60
A
AC
40
20
0
D
0.000
1.019
2.038
3.057
4.076
Effect
Figure 1. Half-Normal Plot of Effects (after square-root transformation)
99
95
Normal % probability
90
80
70
50
30
20
10
5
1
-1.210
-0.550
0.111
0.771
Residual
Figure 2. Normal Plot of Residuals
1.431
12
1.431
Residuals
0.771
0.111
-0.550
-1.210
5.780
9.340
12.900
16.460
20.020
Predicted
Figure 3. Residuals versus Predicted Level of Response
Actual Square root(Avg size)
19.723
17.490
15.257
13.024
10.791
8.558
6.325
E-
E+
Water add
Figure 4. Main Effect Plot for Factor E
13
Actual Square root(Avg size)
19.723
17.490
15.257
C13.024
C-
10.791
8.558
C+
C+
6.325
A-
A+
Interaction of A:DBTDL conc and C:PI Temp
Figure 5. Interpretation Plot for Interaction AC
Actual Square root(Avg size)
19.723
17.490
C-
15.257
13.024
10.791
C-
C+
8.558
C+
6.325
B-
B+
Interaction of B:Acetone/PU and C:PI Temp
Figure 6. Interpretation Plot for Interaction BC
14
Actual Square root(Avg size)
19.723
17.490
15.257
D+
13.024
D10.791
D8.558
D+
6.325
B-
B+
Interaction of B:Acetone/PU and D:Agitation
Figure 7. Interpretation Plot for Interaction BD
260
210
160
110
60
4.50
3.93
4.00
3.33
3.37
Acetone/PU
2.67
Water Addition
2.00
2.80
Figure 8. Hypothetical Response Surface Graph of Polyurethane Particle Size
Fly UP