Engineering Practice

Accelerating Six Sigma Research with the

Definitive Screening Design (DSD) Technique
DSD is a new design-of-experiments (DOE) technique that is expected to b
benefits when using a Six Sigma optimization strategy
Bart Peeters, Marc Roels and Sam Iterative learning process catalyzed
Van Aeken by experimental design
Bayer Crop Science
Data
Guido Desmarets
Stanwick Consultants Data

s
Data

tatistical tools are deeply $ f dsd' %

ingrained in the Six Sigma Screening DOE ’ Full factorial DOE ’ Response Validation
methodology to optimize surface
processes and products ‘ TjtvTmodet.
during chemical process industries 3rd model, >• hypothecs
2nd model, hypothesis
(CPI) operations. In fact, Six Sigma \ st model, hypothesis
has been an important contribut­ hypothesis
ing factor for the widespread use
FIGURE 1. DOE is a useful scientific method to gain knowledge about a particular problem in a manufac­
of statistics in many different indus­ turing process. A classic research strategy follows the sequence of a screening DOE, a full factorial DOE,
trial sectors over the past several and a response surface method to find optimal process settings. A DSD may find these settings with a
decades [7], In Six Sigma's DMAIC significantly reduced number of experimental runs (adapted from [4])
(define-measure-analyze-improve-
control) roadmap, many statistical DOE for scientific investigation a second cycle in the iteration of the
methods are pivotal for the proper When a chemical engineer is con­ deduction-induction process is initi­
collection of data and the abil­ fronted with a manufacturing or ated, and so on [4). Throughout this
ity to translate gathered data into product-related problem, he or she process, the discrepancy between
useful information and actionable will initially propose a first tentative model consequences and data di­
knowledge. In particular, the design model or hypothesis (which is usu­ minishes (this is visualized by the
of experiments (DOE) methodology ally based on speculation) to explainconverging lines, from left to right,
appeals to many chemical engineersor solve the issue, as shown at the in Figure 1), and new knowledge is
(see, for example, Chem. Eng., Nov.left side of Figure 1 [4], From this firstgenerated - the key to process and
2014 and Sept. 2016 issues [2,3]) model, the engineer will then deduceproduct improvement.
as a methodology for systematically certain inherent consequences, Through the different stages of
quantifying cause-and-effect rela­ which, in a next logical step, should experimentation, different types of
tionships between input and output be compared with data to support DOE are typically used, as shown
variables during both manufacturing or refute the first model. Here DOE in Figure 1. In the early stages of
processes and laboratory research comes into play most of the time, research, a screening DOE (called
and development (R&D) efforts. DOEas a technique to acquire new data fractional factorial) is used to identify
is also widely used during the Im­ using a scientific approach. the important input variables (called
prove phase of Six Sigma projects. In many cases, the newly gatheredthe vital few) and to eliminate the ir­
This article discusses the impor­ data will not agree (or will only partially relevant ones (called the trivial many)
tance of DOE in R&D, and the new match) with the consequences of thethat affect the process performance
definitive screening design (DSD) initial model. As a result, engineers or product quality. Starting with a
technique. It also presents a case will usually call for a second, data- large number of potentially important
study to illustrate the power of the driven DOE model, which, likewise, input variables (also called factors),
DSD technique by comparing its will often lead again to some neces­ screening DOEs aim to identify the
results with those obtained from a sary consequences. Some variablesvital few variables that demand fur­
classic DOE process, the latter with from the DOE will turn out to have ther investigation; that is, the active
a much larger number of experimen­or not have an effect on the studied factors that have the largest effect
tal runs. Overall, with this practical problem; some unexpected insights on the response of interest [5],
tutorial, we aim to bring the DSD may appear and the engineer's way In the next stage of research, a
tool to the attention of individuals of thinking about the studied prob­ full factorial DOE consists of all pos­
throughout the CPI. lem will change accordingly. Hence, sible combinations of levels for the
54
CHEMICAL ENGINEERING WWW.CHEMENGONLINE.COM JANUARY 2019
active factors. The purpose here is (-1,1,1)
(0.0. u)
»:><! (-U.D
to quantify the effects of the latter d.1.1) • Center point ♦ Center point
■ Cube point * Edge point
on the response in a more precise (O.a.O) ♦ Star point (1.0.1) ▼ Vertex point
and reliable way using linear mod­ (-xx.O.j) +/ JO.-1.1)
els (main effects and interactions). <1.-1.1 (1.1.0)
# ja. 0.0) (-1. -1.0)
Because the input variables are X3 (. . 1. -1li CQ. 0.0) (0.0.0)
X3
changed simultaneously in a DOE, •W <1.1.-1) JL (0.1.-1)
possible synergistic and antagonistic ./ (0, *a,0)
X2 -1.-0.-1)
X2
interactions between the input vari­ C-1.-1.-1) xi
xt
ables can be detected; this is in con­ * H.-1.-1)
(0. 0, -a) (1.1,-1)
trast to a so-called OVAT approach, FIGURE 2. Shown here is a central composite design (CCD; left side) and a definitive screening design
which involves changing only “one (DSD; right side) for three factors. For the CCD, a is the axial distance of the start points, which are in­
variable at a time.” Equally importantcluded in the experimental design to quantify quadratic (curvature) effects between the response and the
to the detection of possible interac­ input factors
tions is the recognition of eventual TABLE 1. DSD DESIGNS FOR 4 TO 6 FACTORS (ADAPTED FROM REF. 7)
departure from linear relationships N = A N= 5 N= 6
between response and input vari­ run X1 X2 X3 X4 run Xf X2 X3 X4 XS run X1 X2 X3 X4 X5 X6
ables by including a center point in 1 0 1 -1 -1 1 0 1 1 -I -I 1 0 1 -1 -I -1 -1
the design. This is shown on the left 2 0 -1 1 1 2 □ -1 -1 1 t 2 0 -1 1 1 1 1
side of Figure 2. 3 -1 0 -1 1 3 1 0 -1 -1 1 3 1 0 -1 f 1 -1
The first-order design can de­ 4 1 0 I -1 4 -1 0 I 1 -I 4 -1 0 1 -1 -I 1
tect global curvature, but it cannot 5 ■1 -1 0 -1 5 1 ■1 0 1 -1 5 ■I -1 0 1 ■1 ■1
separately estimate the quadratic s 1 a l 6 -1 a -1 1 6 1 1 0 -1 1 1
7 -I 1 0 7 1 -1 1 0 7 -1 1 1 0 1 -1
effects of each factor. At this stage
1 -1 0
of experimentation, interpretation of 8 1 -1 -I 0 -1 -I 8 I -1 -1 0 -1 1
g o o 0 0 9 i 1 1 a g -1 i -i o
the results combined with process -1 -1 -1 -1 10 -1 1 -1 1 1
ID 0 0
and product expertise may allow 11 □ 0 0 □ 0 11 1 1 1 ) -1 0
the identification of the direction of 12 ■1 -1 -1 -1 1 0
steepest improvement toward an 13 0 0 0 0 0 0
optimum in the response, possibly
leading to another series of designedthe engineer to detect curvature in the center values of all variables (This
experiments — this time closer to thethe relation between response and is indicated in red color on the last
optimum — which can yield another the factors. At that moment, star rows of Table 1).
first-order model before proceed­ points (at levels -a -and +a) can The following design pattern fur­
ing to the final stage. Finally, when itbe added to the design afterward ther characterizes a DSD [7):
comes to optimization of the manu­ to allow the engineer to properly 1 .Regarding the location of the zeros
facturing process or product, a morequantify quadratic effects. This flex­ (highlighted in grey color in Table
elaborate model will be needed to ibility makes a CCD very popular in 1) — The first two runs have zeros
describe the region around the op­ industrial process development [6]. in the column of the first variable
timal response and to locate the From the CCD representation shown XI; the next two runs have zeros in
latter. A linear model will no longer in Figure 2, it is clear that the design the column of the second variable
be sufficient. Instead, a quadratic points are uniformly distributed in the X2 and so on
model will be used in a so-called re­ experimental space. 2.Regarding the pair of runs —
sponse surface method (RSM) to fit These are mirrored (folded over),
the optimum. Definitive screening design which means that the second
In contrast to the most familiar run of a pair is found by multi­
Central composite design screening designs where input vari­ plying the first run of this pair by
Since we initially made use of a cen­ables are set at only two levels (-1 -1. Hence, the first 2N runs have
tral composite design (CCD) in our and +1, or low and high level), the exactly one variable at its center
case study discussed below, we will definitive screening design (DSD) value (0), while all other variables
shortly discuss this type of RSM first.introduced in 2011 by Jones and are at their extremes (-1 or +1) and
At the left side of Figure 2, a CCD is Nachtsheim, employs three levels for are referred to as "edge runs,” be­
depicted for three factors. The CCD the variables: -1, 0 and +1; or low, cause the 3-D projections involv­
is very flexible as it can be set up in center and high level (7). For N vari­ ing these variables, they are on the
a modular way: initially, one starts ables the DSD requires only 2/V+1 edges of the cube [8], In case the
with the execution of a two-level experimental runs. In Table 1, the number of factors N is uneven, it is
factorial DOE, where each factor is designs for the case of 4 to 6 fac­ recommended to choose the DSD
set at its low (-1) and high (+1) level tors are shown, as presented in the design for N+1 factors and then
to verify the effects on the response paper from the DSD-inventors Jones to drop the extraneous column,
of interest. A center point, depicted and Nachtsheim. The DSDs are which results then again in an N-
as (0,0,0) in the center of the 3-D comprised of N fold-over pairs plus factor design with 2N+3 runs [8],
representation in Figure 2, may helpone overall center run consisting of The two extra runs are, inherently,

CHEMICAL ENGINEERING WWW.CHEMENGONLINE.COM JANUARY 2019 55

 TABLE 2. DESIGN OF THE RESPONSE SURFACE (90 RUNS)
Run A B C D F F Run A C 0 F F
63 -1 -1 -1 •1 7.29 31 -2.83 0 0 0 0 0 7.92
76 1 -1 ■1 ■1 •1 6.87 78 2.83 0 0 0 0 0 6.72
6 ■1 1 -1 -1 -1 8.37 39 0 -2.83 0 0 0 0 6.40
68 1 1 -1 -1 7.83 42 0 2.83 0 0 0 0 8.43
32 -1 -1 1 -1 -1 6.77 25 0 0 -2.83 0 0 0 8.42
59 1 -1 ■1 ■1 ■1 6.55 26 a 0 0 2.83 0 0 0 6.78
18 -1 1 -1 -1 -1 7.45 17 .. o 0 0 -2.83 0 0 7.19
46 1 1 -1 7.08 64 0 0 0 2.83 0 0 7.21
10 -1 -1 I -1 -1 7.21 44 0 D 0 0 -2.83 0 7.13
22 1 -1 1 -1 -1 6.81 2 0 0 0 0 2.83 0 7.21
1 -1 1 1 -1 -1 8.35 4 0 0 0 0 0 -2.83 7.11
67 1 I 1 -I -1 7,80 35 0 0 0 0 0 2.83 7.29
87 -1 -1 1 -1 ■1 6.84 48 0 0 0 0 0 0 7.26
56 1 -1 1 -1 -1 6.58 40 0 0 0 0 0 0 7.18
12 -1 1 1 -1 -1 7,47 41 0 0 0 0 0 0 7.19
13 1 1 1 -1 -1 7.09 86 0 0 0 0 0 0 7.16
-1 -1 -1 1 -1 7.30 60 0 o o 0 0 0 7.20
74 1 -1 -1 1 -1 6.90 5 0 0 0 0 0 0 7.20
45 -1 1 i i -1 8.23 11 'a a o o o 0 0 7.21
-L
83 1 1 1 1 -1 7.80 27 0 0 0 0 0 0 7.22
80 -1 -1 -1 1 -1 6.81 28 0 0 0 0 0 0 7.17
57 1 -1 1 -1 6.55 79 0 0 0 0 0 0 7.20
38 -1 1 -1 1 -1 7.48 16 0 0 0 0 0 0 7.17
43 1 1 1 -1 7,12 36 0 0 0 0 0 0 7.19
82 -1 -1 1 1 -1 7.25 62 0 0 0 0 0 0 7.17
30 1 -1 1 1 -1 6.85 75 0 0 0 0 0 0 7.23
70 -1 1 1 1 -1 8.32
58 1 1 1 1 -1 7.79 TABLE 3. CONCENTRATION (WT.%) OF VARIABLES A-F
34 -1 -1 1 1 -1 6.85 CORRESPONDING WITH THE CODED VALUES USED IN
66 1 -1 1 1 ■1 6.61 TABLE 2 AND TABLE 4
24 -1 1 1 1 -1 7.47 Coded values
77 1 1 1 1 -1 7.13 -2.83 -1 0 1 2.83
61 ::
dJ
-1 -1 1 -1 1 7.39 A 0449 0.618 0.710 0.802 0971
85 1 -1 ■1 1 6.86 0 0846 1.062 1.180 1.298 1.514
72 -1 1 -1 -1 1 8.39 C 0-702 0.966 1,110 1.254 1.518
84 1 1 -1 -1 1 7.85 D 0430 0.540 0.600 0.660 0770
37 -1 -1 -1 -1 1 6.86 1 3.585 4.500 5.000 5.500 6 415
49 1 -1 -1 1 6.56 f 0.108 0.135 0.150 0.165 0.192
29 -1 1 -1 1 7.49
50 1 1 TABLE 4. DESIGN OF THE DSD (13 RUNS)
rn

9 -1 -1
CD

Si

21 1 -1
O

Q
i kn co bo u a I ro
AJ 05 O) N CO O) N|N

CD
5'.
Q

-1 1 -'■I
*V|

: a
19 1 1
oi

-
65 -1 -1
CD
ro

Edge point

1
<D

52 -1
—j
*Ail(o|ul(o|co|ui|coco4>.knct>uio>|-u

-ts.

71 -1 1
O
-----

8 1 1
SG

3 -1 -1
rc
o>
CO
CO 05 AJ

CO

47 1 -1
CO
CO

o
ioiw

-i i

69 -1 1
—-in
co co w

CTJ

73 1 1
O

T
CO CO

20 -1 -1
CO

CO

in:

54 1 -1
i\>

in
o

a
O
it oi

16 -1 1
23 1 1
07
co I oo co n co to in'

89 -1 -1
' cd oo N cb ib nIs

53 1 -1
55 -1 1
,N 9 CO S O sT (O

14 1 1
81 -1 -1
90 1 -1
33 -1 1
7 1 1

56 CHEMICAL ENGINEERING WWW.CHEMENGONLINE.COM JANUARY 2019

 CO

w 5>
o

a —n
u

2? ©

</s

a
UJ iuj

X CO

o
S
ca

UJ

c/> o
in u
c/> m

m m
o

5)5
CO

tn
8£
Oa
si

•ea
=r O
CM

\
■

U
LJ
ii

Oh
a
\

03
S
\
\

O
CO

\
1'S
si
§§

-3S,
E E
=r o

\
i

■\

II
II
*
<JH

\
Predicted pH

-
\

cn

X
\

CO

co
+

CJl

X
3G M

X
\

:T:

X
V\

+
CM

i \ •'

6
cn
i

00
X

X
N

--
x x

h-
cc

O
■T.

X
'.1
----
a>
\

+ +
O

Is
9
NXt

OQ
x x
CM
03

ZJ>
6.50 MODEL SUMMARY
6.50 6 75 7.00 7.25 7.50 7.75 8.00 8.25 850
Measure pH
R! 0.996 0.991
FIGURE 3. Predicted pH values of 78 mixtures (from the CCD experiments)
0.995 0.985
using the DSD model, and the predicted pH values of 13 mixtures (from the^ad| ;i'l|

DSD experiments) using the CCD model, to validate both models. n-l«"l 0.992 ;r l-.-l 0.977
without center values for any fac­ DSD design for N = 4 from Table 1 sive technique for users in the field
tor and are referred to as “vertex and then dropped the last column. 2. Ad vantages are specifically found
runs” by the DSD-inventors [8]. One can clearly see the so-called with regard to confounding vari­
Likewise, to increase the power of DSD edge runs and vertex runs ables (note: The term confound­
any DSD, one may choose initially (which comes down to cube points), ing, used in statistics, indicates
a design with a larger number of together with the single DSD center that the effects of model terms
factors than necessary (for exam­ run at the right side of Figure 2. cannot be calculated separately):
ple, in total N+ k factors), and then The advantages of a DSD are as • The main effects are completely
drop the extra k columns [7-9]. follows (For more, see (7-9]): independent of each other, com­
For the sake of visualization, a 3-D I.Each factor is analyzed at three pletely independent of two-factor
representation for a DSD with three levels, which makes it possible interactions, and completely inde­
factors is shown at the right side of to analyze quadratic effects, and, pendent of quadratic effects (thus,
Figure 2, although a DSD with only thus, to model curvature with a there is no confounding at all)
three factors is not recommended very limited amount of runs, ren­ • The two-factor interactions are not
[9]. For this visualization, we took the dering DSDs a relatively inexpen­ completely confounded with other

Hot Topics 1
in the Managing Vapor and
1

Particulate Emissions

Chemical Valves Guidebook

Vi Optimal Pump
Processing •'A
Management

Industry [viJrwnWfF ■ ms m

Valves Selection: Optimal Pump Managing Vapor

o CHEMICAL
ENGINEERING Operation and Management
Troubleshooting
and Particulate
Emissions

Find these and other related CPI titles in the store.chemengonline.com

CHEMICAL ENGINEERING WWW.CHEMENGONLINE.COM JANUARY 2019 57

 two-factor interactions, although tions (wt.%) of the variables A-F are the case of the CCD, this would be
they may be correlated a little summarized in Table 3. The mea­ 90-1 =89 observations). Then the
• The quadratic effects are esti­ sured pH of the formulations can be pH is calculated for the mixture that
mable and are independent of found in the last column. was intentionally ignored to build the
the main effects (no confounding) After these CCD experiments model. Next, the predicted residual
and not completely confounded, were executed, Six Sigma training error is estimated for this single ob­
though correlated, with the two- was given at the Bayer Antwerp site servation; that is the difference be­
factor interaction effects (where the work was carried out), in­tween the calculated pH (pHca/c)
• Last but not least, for DSDs with atcluding the new statistical technique and the experimentally determined
least six factors, the DSD can fit —of DSD. In the context of the train­ pH (pHobs) for the observation. This
with very high level of statistical ef­ing, the Six Sigma project leader procedure is repeated for all obser­
ficiency — the full quadratic models sought to verify which model would vations. Finally, the so-called pre­
in case only three or fewer factors be found using a DSD with only 13 dicted residual error sum of squares
turn out to be active. As such, the runs for the six factors, compared to (PRESS) is calculated according to
DSDs become efficient response the CCD model obtained earlier. For Equation (1):
surface designs with three or fewerthis, 13 extra formulations were pre­
factors, rendering follow-up experi­pared, as shown in Table 4. PRESS=
ments (for the purpose of model The data sets were analyzed using 0)
optimizing) not necessary in many Design-Expert (from Stat-Ease Inc.),
circumstances [7]. “The capability including only model terms at a sig­ This PRESS value is compared to
to project these designs to efficientnificance level of 0.05. A summary ofthe Sum of Squares (SS) around the
response surface designs makes both models is depicted in Table 5. mean value of the whole data set
possible the screening and optimi­The CCD model includes the main (SStot); in other words, in the latter
zation of a system in a single step,”effects A, B, C and F, the interactioncase, a “mean model” is used that
according to the co-inventors effects AB, AC and BC, and the qua­simply takes the mean response as
Jones and Nachtsheim [8]. This dratic effects A2, B2 and C2. The DSD a prediction for every mixture in the
potentially huge benefit (making a model includes fewer terms, namely data set. The fl2pre<J is finally found
shortcut from screening straight tothe main effects A, B and C, and thevia Equation (2):
optimization) is shown in Figure 1 interaction effects AB and BC. It is
with the dashed purple arrow. For remarkable that the coefficients in PRESS
comparison, a response surface the DSD model are very similar to the R2l>i< </ = 1-
A.S
design with three factors would re­one of the CCD model. (2)
quire 20 runs. Regarding the CCD model, the
In the original DSDs, the factors hadR2adj (that is, the raw coefficient of The somewhat larger Fhpred of 99.2%
to be quantitative to allow for 3 lev­ determination, R2, being adjusted forfor the CCD compared to 97.7%
els. New DSDs have very recently the number of model predictors) in­ for the DSD model suggests that
been developed that allow for com­ dicates that 99.5% of the variation inthe CCD has some greater predic­
binations of continuous and categor­the pH is explained by the variation tive ability. However, the DSD model
ical input variables [70]. in the input variables; based on the as such can certainly be classified
complete CCD statistical analysis as a model having a great
Case study (not shown) the terms F, AC. A2, B1 predictive power.
Part of a particular Six Sigma projectand C2 — which are not included in Another way (and by far a better
consisted of making a model using athe DSD model — account only for way) to validate a model is to use a
CCD for the pH of formulations made2% of the explained pH variation in completely new data set that was
of six ingredients (A-F), whereby thethe data set of the CCD, which is, not used during the development
ingredients were varied indepen­ after all, a rather minor contribution of the model. This new data set —
dently of each other. The number of these extra CCD model terms. used in the stage of model validation
of laboratory test formulations pre­ Based on their high R2ad| (99.5% — is frequently called a test set, in
pared and evaluated was 90. In a for the CCD, and 98.5% for the contrast to a training set that is used
laboratory-scale environment, this isDSD) we can state that they each to calibrate (build) the model [7 7],
doable. However, in a manufactur­ fit well with their respective existing The test set should ideally span the
ing plant, this would be difficult, per­ data set. same “space" of the input variables
haps impossible. The different test But how do both models compare as was the case for the training set.
runs are shown in Table 2, where thewhen it would come down to pre­ We then let the calibrated model
cube points, star points and center dict the pH of new mixtures? We will predict the responses of the test set
points are clustered together for the first have a look to their predicted R2and compare them to the known,
sake of overview. The run order as iffpred)- This model summary is cal- real responses.
executed in the laboratory is shown culated by setting aside a single ob­ In this case study, we will use the
in the first column. The variables areservation from the whole model datadata set of the CCD as a test set for
shown with their coded levels, with set, and then re-estimating the modelthe DSD model; likewise, we will use
the value of a (star points) being based on all observations minus the the DSD data as a test set to validate
2.83. The corresponding concentra- one that is excluded (for example, inthe CCD model. Because the 12
58 CHEMICAL ENGINEERING WWW.CHEMENGONLINE.COM JANUARY 2019
 axial points in the CCD data set are The DSD methodology lately de­ He holds a master degree in chemistry (University of
Ghent 1976). Subsequent research work resulted in
beyond the design space, wherein veloped by Jones and Nachtsheim several publications and patents on slow-release for­
the DSD model was built (remem­ is highly efficient with the number mulations for pesticides.
ber, the DSD model factor settings of runs required far below the num­ Marc Roels is a crop protection
were between -1 and +1, whereas ber needed in classical screening products laboratory specialist at
Bayer Crop Science Europe N.V.
the CCD star points were going fromdesigns. As a low number of ex­ (Phone +32 3 568 5185; Email:
-2.83 to +2.83), we will exclude the perimental runs is in most cases a marc.roels@bayer.com), where
CCD mixtures with the star points to desirable requirement for many ex­ he has been working since 1984.
Roels provides analytical support
validate the DSD model; hence, the perimenters in the CPI, and certainly tor research, development, regis­
pH of the remaining 78 CCD mix­ in cases where experiments need tration and manufacturing of her­
A.
tures were predicted with the DSD to be done in manufacturing plants, bicides at the Bayer Antwerp site
model and compared with the mea­ the DSD can be called a revolution and at different toller operations. For more than 15
years he is an enthusiastic user ol DOE to study formu-
sured pH values. Figure 3 presents in performing designed experiments.lation robustness and formulation optimization. He re­
the predicted versus measured pH In case the statistical analysis of ceived his 8.S. in pharmaceutical and biological tech­
values for the validation of both mod­the experiments, a fortiori indicates niques in 1983. He is a certified Si* Sigma Green Belt.

a
els. Visually, one can see that both that the number of active factors is Sam Van Aeken was the team
lead of the analytical services
models adequately predict the pH, limited, so for these cases, the use laboratory at Bayer Crop Science
with some more variation in the caseof definitive screening allows the (Haven 627, Scheldelaan 460,
of the DSD model. In the context engineer to make a shortcut from 2040 Antwerp. Belgium), where
of model validation, the root mean screening straight to optimization. he has been working since 2014.

aTI
Prior to joining the firm, he worked
square error of prediction (RMSEP) It can be expected that the CPI will from 2010 until 2012. as a pro­
can be calculated as an estimate of benefit from this latest statistical re­ cess chemist in Eastman Ghent,
the prediction error [11], using Equa­search, and that definitive screening managing new product trials. He
was later employed at W.R. Grace / DeNeef as a R&0
tion (3): designs will be applied with growing engineer, where he researched new urethane based
frequency over time, especially in grouting systems. Van Aeken graduated as a bio-engi­
CPI Six Sigma projects where a lot neer in 2004 at the Vrije Universiteit Brussel and ob­
PH.J- tained a Ph.D, in organic chemistry in 2010 with a dis­
of factors need to be tested. ■ sertation
RMSEP - i

on new synthetic routes towards

N (3) Edited by Suzanne Shelley
aza-heterocyclic quinone compounds.

The RMSEP can be interpreted as

the average error to be associated Authors References
with future predictions. In practice, Bart Peelers is a manufacturing1. Steinberg. DM.. Industrial statistics: the cbalenges and
f0.025; n-1 times the RMSEP (with technologist at Bayer Crop Sci­the research, Quality Engineering, 28 (1), pg 45-59,
ence (Haven 627, Scheldelaan 2016.
t obtained from a Student’s t table 460. 2040 Antwerp. Belgium; 2. Kleppmann. W.. Design ol Experiments (DOE): Optimiz­
[12], and n the number of data Phone: +32 3 568 5762; Email:ing products and processes elficientty, Ctem. Eng., No­
points of the test set; for example, bart.peeters@bayer.com). where vember, pp, 50-57,2014,
based on the test set to validate the he has been working since 1998.
3. Anderson. M.J., Design of Experiments (DOE): How to
He first served as a process im­
CCD model, foo25' 13-1 is 2.16. With provement engineer at Eastman's handle hard-to-change factors using a split plot, Chem.
jA Eng. September, pp. 83-86,2016.
larger test sets, the t value will ap­ PVB polymer manufacturing plant
proach 1.96. This may be used as onsite (until 2004). Since then, he has been working 4. Bo*. atG.E.P.. Hunter. JS. and Hixiter, W.G. "Statistics for
estimated precision for predicted the environmental department of the company. experimenters:
is a certified Six Sigma Black Belt and coordinates
Peeters
ed.. the
design, innovation, and discovery". 2nd
John Wiley & Sons, Inc. 2005.
pH values. For the CCD model the Si* Sigma program at the Bayer Antwerp 5. Jonessite. While
B. 21st century screening experiments: what,
RMSEP is 0.07 pH units, and for theworking at Bayer's WWTP, he obtained his how.
why and Ph.D.Quality
in en­ Engineering 28 (1), pp. 98-106,
gineering from the KU Leuven (Belgium) on the2016 research
DSD model, it amounts to a some­ topic "Effect of activated sludge composition on its de-
what higher value of 0.09 pH units. waterability and sticky phase.'Prior to that, he 6.earned
Anderson, M.J "RSM simplified: optimong processes
For example, the pH of the mixture an M.S.Ch.E. degree from the KU Leuven, and using surface methods tor design of expenments", 2nd
an M.Bio.
of DSD run 13 with an observed Ch.E. degree in 1996 from the university collegeEd., DeProductivity Press 2016.
Nayer. Peeters is the author of 20 articles in scientific
pH of 7.40 (first row in Table 4) is journals, technical magazines and international7.desings
Jones B. and Nachtsheim, GJ A class of three-level
confer­ for definitive screening in the presence of sec­
predicted by the CCD model to be ences. ond-order effects Journal ol Quality Technology 43. pp.
7.30 +/- 0.15 (the latter being 2.16 Guido Desmarets is a senior 1-15.2011.
times 0.07). management consultant at Stan-
% wick 8. Jones B and Nachtsheim, CJ Blocking schemes for
NV (Axess Business park­definitive screening designs Technometrics 58 (1), pp.
building B-Guldensporenpark 20. 74-83,2016.
Final thoughts 9820 Merelbeke, Belgium; Phone
>> +32 9 210 59 50: Email: guido. 9. Jones B, JMP Blog on proper and improper use of
Chemical engineers and scien­ Definitive Screering Designs (DSDs) (https://community.
tists in the CPI embrace the DOE desmarels@stanwick.be), where jmp.com/t57JMP-Blog/Proper-and-improper-iise-of-De-
he has been working since 1988, fimtwe-Screenmg-Desigrts-DSDVba-p/30703) 2016.
methodology developed by statisti­ As a Master Black Belt he coaches
cians to efficiently investigate and worldwide companies in the field10. Jones B. and Nachtsheim, GJ, Definitive screening
optimize products and manufactur­ of Continuous Improvement, Lean Six Sigma, Opera­ designs with added two-level categorical factors Journal
tional Excellence, with major results in efficiencyotand
Quality TechnologtA5, pp. 121-129.2013
ing processes. We therefore grate­ quality improvements, cost reductions and delivery
11. per­ Esbensen. K.H. "Multivariate Data Analysis - m Practice
fully acknowledge the work done formance. Prior to joining Stanwick NV he held An posi­
Introduction to Multivariate Data Analysis and Experi­
by statisticians to constantly search tions as R8D manager, process engineer, quality assur­
mental Design". 5th Ed.. CAMO software. 2010.
for even more efficient design-of- ance and production management at international
companies, where his interest to use DOE as 12. Bass, l. "Six Sigma Statistics with Excel and Mntab",
a major
experiments approaches. 1st Ed., Me Graw Hill, 2007.
process improvement tool was triggered and developed.

60 CHEMICAL ENGINEERING WWW.CHEMENGONLINE.COM JANUARY 2019