Compound Development With Design of Experiments or The Grafcompounder?

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 9

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/281493253

Compound Development with Design of Experiments or the


GrafCompounder?

Article · August 2013

CITATIONS READS

2 745

1 author:

Hans-Joachim Graf
H-JG Consulting
50 PUBLICATIONS   147 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

compound simulation View project

Injection Molding Process and Lifetime of a NR-Rubber part in dynamic application View project

All content following this page was uploaded by Hans-Joachim Graf on 05 September 2015.

The user has requested enhancement of the downloaded file.


Special Reprint from
Magazine for the Polymer Industry RFP Rubber Fibres Plastics

Compound development with


design of experiments or
the GrafCompounder?

Volume 8, August 2013

03| 2013
Computation of compounds

Compound development with design of pany, but hidden to him. With the amazing
number of recipes generated within a rubber

experiments or the GrafCompounder? company it is quite possible that searching


the archived recipes carefully would reveal
the newly created recipe being just a reinven-
H.-J. Graf* tion of the wheel. This is even more likely in
companies without standardisation of ingre-
dients for avoiding redundant raw materials.
The analysis of compounds in a database is limited to functions such as search-
Many chemists do not realise that often very
ing and identifying compounds. Programmes calculating formulations by applying
mathematical models of neural networks failed to predict a recipe with sufficient similar recipes are created, only differing in
accuracy. A programme using multiple linear iteration procedures can signifi- brand names of the ingredients used. Even
cantly optimise the accuracy of the prediction because it allows the identification if more advanced compounders would uti-
and elimination of corrupted data within the database. Such a method improves lise systematic approaches such as statistical
the data quality considerably. For checking the accuracy of the prediction two experimental design (DoE), some value will
examples have been used. For this exercise data sets from a statistical experimen- be lost. Typically, almost all the recipes used
tal design have been taken and the calculated results obtained by the prediction in a DoE are ignored or discarded when the
tool of the DoE programme as well as the programme GrafCompounder have been
optimum recipe has been found with that
compared to each other. It became apparent that the variance of the prediction
results of both programmes fell into the 95 % confidence interval. technique and the confirmation has been
successfully done. If an adjustment is still
necessary, the compounder will use the rec-
ipe as a starting point. The compounds from
1. Introduction pounds to manage process improvements, to the DoE are already archived meaning out of
increase process safety or to reduce manu- sight. This leads us to the question, how we
Many companies develop their own rub- facturing costs. can extract and utilise the value stored in old
ber compounds in house, typically building recipes, compound databases, or archives for
a large database of recipes over years of A typical rubber company is in operation future compound developments in order to
operation. These recipes naturally contain for several decades. It should not be a sur- save valuable time and money.
a set of ingredients and a set of properties. prise, that a company with a diverse rub-
ber article range has some 500 – 1 000 living In the recent past there were some inter-
New rubber formulations are often recipes (not forgetting the demand for spare esting approaches dealing with this problem.
created with varying methods. The method parts) and tenfold the number of laboratory In the literature some calculation methods
being mostly used is trial and error, common recipes taking into account the number of based on non linear regression were pub-
is still a one step at a time method and there recipes created during small scale evalua- lished [1 – 5] and a tool based on mathemati-
is an increasing number of statistical models tions. It can be reasonably assumed that all cal models of neural networks was available
for the design of experiments (DoE). these compounds have been manufactured on the market [6]. Extensive tests were done
at least once in a small laboratory mixer and with this programme, but predictions with
The development of a new compound can that their physical properties have been test- sufficient precision failed even with a large
originate from the demand for improve- ed too. We have to consider, that the cost database created and used for the query.
ment, validation of raw material properties, for mixing and testing of a single compound
or quality concerns. New recipes are often is worth approximately USD 500. This means The following is a description of a pro-
created based on existing compound for- that such a data base represents a substan- gramme having been developed recently. It
mulations for meeting new specifications, tial financial investment. But how often do applies a concept of the compositional ratio of
due to changing customer demands, or to we use this significant investment to our ad- a multi component material being determined
improve deficient products that suffer ser- vantage? What do we get in return? automatically by using a multiple linear itera-
vice life failures. New compound recipes are tion method (MLI) which is based on the linear
even created during the fine tuning of com- Whether a compounder takes a starting relationship between the material’s compo-
recipe from the literature of a material sup- nents and their physical properties [7, 8 – 12].
plier or uses an existing production formula-
tion, the industry standard is to create some
* Hans-Joachim Graf, recipes in the laboratory, always performing 2. Motivation for developing
hjgraf@cyg.net laboratory trials and evaluations followed the programme
Rubber Industry Consultant by plant trials. This procedure increases the
HJG Consulting, Bad Soden-Salmünster, number of recipes. Most likely a compounder The initial motivation for designing a new
Germany will not check whether his new compound is compound development programme was not
already available in the database of his com- prompted by the desire to recapture value

2 RFP 3/2013 – Volume 8


from previously developed formulations. exhibited errors, the mathematics did not al- programme has used for the calculation of
Rather it emerged due to lack of success low smoothing of the data compared to a the new compound. As a precondition the
when working with a commercially available multiple regression analysis [13]. Since there programme should run on all personal com-
compound development programme based was no tool available for identifying faulty puters no matter which operating system
on a non linear approach like neural network data sets there was no chance to clean the is in use. Data transfer/conversion should
mathematics – as said above. The aim was to database. be possible via copy/paste with almost all
reduce the amount of trial and error when standard spreadsheet software.
formulating and to foster a more scientific The idea of building a database only with
approach to the creation of recipes. How- compounds made within a statistic design of The answer to these objectives is a pro-
ever, after more than one year of testing, it experiment (DoE) would not eliminate the gramme called GrafCompounder. This pro-
was observed that the programme repeatedly problem, because the database entry would gramme was extensively tested by the author
failed to predict accurately the properties of be the property as measured but not cor- in all kinds of computer environments and
the compound created. The database in use rected via regression. Man is able to iden- under all kinds of compounding challenges
was composed of compounds utilised on a tify a property as an outlier in his analysis like simple optimisation of physical proper-
large scale and compounds created in a lab- of variance and may repeat the experiment ties up to more difficult tasks like surface
oratory environment at the same time. We putting a new set of data in his database appearance of compounds (fig. 2).
found out, that we were unable to predict a for correction. The DoE programme, however,
compound with its properties to match the will always use data fitted with its regression
target property values accurately enough, equation just ignoring outliers. 3. Testing the GrafCompounder
which means within the 95 % confidence
interval. We concluded, that the database If a multi component material with a plu- To show the ability of the GrafCompounder
contained too many faulty data. Just to give rality of physical properties must be inves- to predict a compound from any given data-
an example: The tensile strength of some tigated and if linear relationships between base, a comparison with DoE data was cho-
compounds targeting at a high tensile was incorporated ingredients and corresponding sen. The DoE programme uses a prediction
much too low compared to the predicted val- properties can be implied the analysis and and optimisation tool to calculate a com-
ues [6]. The confirmation experiments failed. optimisation could be handled by a pro- pound. The result of this calculation can be
The mathematics of the programme seemed gramme much easier. Using linear relation- used for the purpose of comparing both cal-
to be intolerant regarding such faulty data. ships should make the design of a material culation methods. As DoE programme De-
The programme – as we would say - could composed of many components possible sign-Expert was selected.
not learn, ignore or eliminate doubtful with sufficient precision.
data. If the ingredient property relationship The difference between both tools:
Looking into the old compounding tables Regression on the one hand, iteration on
Fig. 1: Design guideline for developing the pro- published in the past [7 – 12] you will find in the other, has some influence on the result
gramme GrafCompounder most cases a linear relationship for almost all of course. The DoE tool uses statistical model
• Calculation with linear relationships
basic physical properties. Other publications equations for data fitting and analysis within
 Most DoE show a linear model equation to be are showing that there is a linear relationship the experimental design limits. It calculates
sufficiently accurate.
 Math should be based on linear relationships, found with the exemption of the tensile at an area in between the boundaries. Inside
but allow multiple small steps during calculations. break. The latter could be true, because the this area the properties are met in between
• Identification of faulty data in the compound database ingredients were varied within a much larger given limits. If all targets are matched the
should be easy.
• Programme should work correctly even with a smaller
range, which may cause such a non linear- probability is equal to 1, if not all targets
database. ity. We still can assume, that if changes of are matched the probability is less than 1.
• Programme should be able to exchange data with all ingredients are done in smaller steps, a linear
types of calculation programme.
relationship will sufficiently describe the de- The tool GrafCompounder works with dif-
pendency of properties on ingredients. Hav- ferent starting points, which are given by
Fig. 2: Features of the calculation software Graf- ing such a programme tool one would allow the compound data in the database. It will
Compounder a compounder to learn a lot about the real blend mixes mathematically and calculate
value of his compound database. the properties as arithmetic average. A rat-
The GrafCompounder is a:
• Spreadsheet software ing function estimates the distance from all
 based on Java Considering what has been mentioned targets. Using multiple linear iteration and
 import / export function for communication above a design guideline was created and minimisation of the difference between the
 allows automatic mixing of compounds and manual handed over to a programme developer actual properties achieved and the desired
mixing
(fig. 1). The design guideline consists of few properties the targets are reached (fig. 3). As
 calculates profile of properties
requirements to meet the objectives of sim- soon as all the targets are met – that means
 shows data composition of the results
 import / export of results plicity, sufficient accuracy and ease of iden- the rating function equals zero - the Graf-
tifying the compounds in the database the Compounder yields one compound with its

RFP 3/2013 – Volume 8 3


Computation of compounds

properties. In case of a plurality of bounda- Now using the point prediction tool im- Ohm and Rajan Vara [14] has been analysed
ries there are most likely conflicting targets. plemented in Design-Expert we can set the using both tools. This DoE is dealing with
In this case the rating function will show a factors to the same values given by the Graf- 7 factors and needs 41 experiments to be
number different from zero with no further Compounder, analyse the responses, and conducted resulting in a little larger data-
decrease, indicating, that there is no better check, whether they are within the 95 % base of compounds than the example before.
solution possible. Regarding the purpose of confidence interval. Table 3 illustrates that
this comparison the question is, whether this the results of both software systems are in- The correlation of tensile and elongation
single point is inside the area of solutions side the confidence interval and within the at break is linear as expected. Measured and
obtained by the DoE programme. At least it measurement errors of both methods. predicted data (fig. 5) indicate that there
should be inside the confidence interval of must be a difference between the compound
the area of solutions. As a first conclusion it is obvious that a DoE calculated by the optimisation tool of De-
programme, which is using a set of data and sign-Expert and by the GrafCompounder. The
3.1 Comparing DoE and GrafCompounder calculating the responses by regression model
by an oil/filler optimisation equations – either linear or non linear – will Tab. 1: Oil/filler experiment – fractional factorial
output fitted and smoothed data. Calculations design
As a first example a simple oil/filler DoE with the fitted data will yield results, which
Factor Name Unit Min Max
was chosen, performed as a fractional facto- should be somehow different compared to a
rial design, to show basically how these tools software applying a stepwise iteration pro- A CB 6630 phr 60 95

react on a limited set of data, which can eas- cess without any data fitting. Our expectation B CaCO3 phr 10 70

ily be overviewed by an experienced com- would be that there should be a characteristic C Clay phr 70 95

pounder. The factors of this DoE are shown difference but inside a small tolerance, which D Oil phr 70 95

in table 1. is indeed what we observed.


Tab. 2: Oil/filler experiment – conditions selected
Eleven experiments are necessary to per- 3.2 Comparing DoE and GrafCompounder for the optimisation
form the DoE with three replicates of the by a polymer/accelerator optimisa-
Response Limit
centre point. As responses for evaluation tion
the Mooney viscosity (MV), the scorch time Mooney viscosity MV 125 35 – 36 MU

at 125 °C (t5 – 125 °C) and the scorch time In the second comparison of a DoE and the Scorch time 125 °C t5 3.9 – 4.1 min

at 175 °C (t10 – 175 °C) were chosen. All GrafCompounder a DoE published by Robert Scorch time 175 °C t10 0.43 – 0.44 min

responses are significant using a linear re-


gression equation. The conditions selected
for the optimisation are given by table 2.

The optimisation with the software


Design-Expert software showed somehow Path 3

different results for the factors compared Path 4


to the GrafCompounder (fig. 4). The main
differences are the amounts of CB 6630
and CaCO3. To understand the difference
an additional condition was implemented
in the query of the GrafCompounder: The Path 2
amount of CB 6630 was set at 73 phr simi-
lar to the result of the optimisation with Path 5
Design-Expert. Now the results of the two
methods were much closer, the GrafCom-
Path 1
pounder had little less CaCO3 and clay for
its suggested compound. Are this differ-
ences important regarding the responses?
Analysis based on Report of results
To get an answer we used the function • measured data • recipe
point prediction of the software Design- • targets • all calculted physical
• weightings properties
Expert which is nothing else than a com- • rating functions show the distance  missing data ignored
pounding tool similar to the GrafCom- Fig. 3: between values and targets • show all recipes with their
Description of the mul- • iteration in small steps from different percentage used in an analysis
pounder but based on regression math- starting points
tiple linear iteration
ematics and limited to the set of data used • check of maximum agreement with
method applied by the the target
in the design. GrafCompounder

4 RFP 3/2013 – Volume 8


reason is the difference between measured
and fitted data, which is shown in the cor-
Ingredients Unit DoE Optimisation GrafCompounder
relation graph.
CB 6630 phr 73 73
CaCO3 phr 68 61 In case of a DoE programme the calcula-
Clay phr 39 32 tion of compound properties will use fitted
Paraffinic oil phr 72 70 data, which do not deviate from the corre-
MV 125 MU 34 34.1 lation curve. The MLI method, however, will
t5 (125 °C) min 4.04 4.1 use the raw data instead. Knowing that the
t10 (175 °C) min 0.45 0.45 prediction of a compound by the DoE pro-
gramme compared to the result of a con-
Overlay Plot
70.00 firmation experiment is inside a 95 % con-
 fidence interval, one would expect, that a
64.00 MV: 34.008
 t5: 4.032 compound prediction by the MLI method
58.00 t10: 0.436 t5: 3.902
X1 72.16
using the GrafCompounder should also give
52.00 X2 60.84 a result within the 95 % interval.
t5: 4.100
46.00
t10: 0.439
B: CaCO3

MV: 34.300 MV: 36.000 As responses for the optimisation were


40.00 t10: 0.435
chosen: hardness from 65 Shore A to
34.00 70 Shore A, tensile from 11 MPa to 12 MPa,
Fig. 4: and elongation from 350 % to 400 %. The
28.00
Comparison of DoE opti-
misation and GrafCom- DoE optimisation within these bounda-
22.00

Optimisation area calculated with Design Expert
pounder quest results ries resulted in a polymer blend contain-
16.00 
Solution given by GrafCompounder with - top: table of results, ing 5.6 % ENB, 0.44 phr sulphur, 2.11 phr
the additional condition (CB 6630 = 73 phr) bottom: optimisation
10.00 DTDC, 1.00 phr MBT, 1.50 phr TBTD, 1.50 phr
graph with flag indicat-
60.00 67.00 74.00 81.00 88.00 95.00 ZDBC, and 1.50 phr DTP. The compound
A: CB 6630
ing responses from Graf-
Compounder properties of this composition were: ten-
sile = 11.2 MPa, elongation = 335 %, and
Tab. 3: DoE optimisation – comparison of the results by GrafCompounder and DoE point prediction hardness = 66.5 Shore A, which is indicated
by the flag in figure 6. Using the point pre-
Ingredients Unit DoE Optimisation GrafCompounder DoE Point Prediction
diction method instead of graphical optimi-
CB 6630 phr 73 73 73 sation yielded the same ingredient amounts
CaCO3 phr 68 61 61 as before.
Clay phr 39 32 32
Paraffinic oil phr 72 70 70 Keeping in mind the difference of the
MV 125 MU 34 34.1 34.2 ± 3 calculation methods in both programmes
t5 (125 °C) min 4.04 4.1 4.01 ± 0.25 the boundary conditions now were chosen
t10 (175 °C) min min 0.45 0.45 0.43 ±0.07 in a way that one could expect the same
values as indicated by the flag. The lower
Fig. 5: Correlation graph of predicted and measured data for tensile strength and elongation limits of the responses were set at exactly
these values for calculation: tensile from
16.00 Tensile 600.00 Elongation
11 MPa to 12 MPa, elongation from 350 %
14.00 500.00 to 400 %, and hardness from 65 Shore A
to 70 Shore A. After processing the data
Predicted

Predicted

12.00 400.00
with the GrafCompounder slightly differ-
10.00 300.00 ent results were obtained as shown in fig-
8.00 200.00 ure 7. The main differences observed were
the higher levels of ENB and sulphur and
6.00 100.00
a lower DTDC level to achieve the targeted
6.00 8.00 10.00 12.00 14.00 16.00 100.00 200.00 300.00 400.00 500.00 600.00
Actual Actual properties.
 Tensile at break is significant with linear model
• Sulphur has larger influence followed by DTDC and TBTD If the data were processed by the DoE
 Elongation is significant with quadratic model, but linear model is a sufficient fit
• Sulphur has the largest influence followed by DTDC programme with the same boundaries as in
 Hardness is sufficient significant with linear model as well the GrafCompounder there occurred a shift
• Main influence sulphur, DTDC
of the ingredient levels as displayed in fig-

RFP 3/2013 – Volume 8 5


Computation of compounds

ure 8. The results are indicated by the yel- limits for calculation. For achieving the same very different approaches do not differ from
low field (result DoE optimisation) and by the targeted property values the DoE still sug- each other in reality at all.
flag (result GrafCompounder). But what is of gested a lower amount of sulphur and a little
most interest: the sulphur level rise within lower ENB level (fig. 9).
the DoE result while the ENB level is similar 4. Conclusion
to the value before and the DTDC level re- Anyhow, an experienced compounder will
mains high. notice, that a confirmation experiment in A software using a multiple linear itera-
the laboratory or even a large-scale test at tion method is able to predict a compound
How close both solutions really are could the plant will not reveal any differences out- from unorganised data in any compound
be demonstrated by using point prediction, side the measurement error and the 95 % database. Up to now such data could not
which means – as already mentioned above confidence interval for the physical proper- be utilised except for the selection or iden-
– that the values calculated by the GrafCom- ties of both compounds. This leads us to the tification of a compound as a starting point
pounder were put into the DoE software as second conclusion. The results of the two for further development work. The reason

Overlay Plot Fig. 7: Analysis by the GrafCompounder - results


1.50 and boundaries
  Factor values giving this results
• ENB: 5.58 %
• Sulphur – 0.44 phr Ingredients Result
• DTDC – 2.11 phr
• MBT – 1.00 phr A: ENB 6.5
1.20 • TBTD – 1.50 phr
• ZDBC – 1.50 phr B: DTDC 0.98
• DTP – 1.50 phr C: Sulphur 0.93
Tensile strength: 12.000
D: MBT 1
C: Sulphur

0.90
E: TBTD 1.51
F: ZDBC 1.33
G: DTP 1.45
Tensile strength: 11,214
Elongation at break: 335,106 Results
Hardness: 66,470 Elongation at break: 300.000 Tensile strength 11.5
0.60
X1 5,58
X2 0,44 Elongation at break 325
Tensile strength: 11.000
Hardness: 67.000 Hardness 67
Elongation at break: 350.000 Fig. 6: 
Boundaries
0.30
Graphical optimisation • Tensile strength-MPa: 11.5 – 12.0
5.00 6.00 7.00 8.00 9.00 • Elongation at break-%: 325 – 335
A: ENB
by DoE – factors indi-
• H-Shore A: 65 – 67
cated by flag

Fig. 8: Reprocessed data with the same boundary conditions for both software pro- Fig. 9: Comparison of predictions by GrafCompounder and Design-Expert for the
grammes same target properties
Overlay Plot
1.50
Ingredients Result GrafCompounder Result Design-Expert

A: ENB 6.5 5.45


Tensile strength: 12.000
Tensile strength: 11.536 C: Sulphur 0.93 0.88
1.20 Elongation at break: 306.017 B: DTDC 0.98 0.98
Hardness: 68.146
X1 6.50 D: MBT 1 1
X2 0.98 E: TBTD 1.51 1.51
Tensile strength: 11.498
F: ZDBC 1.33 1.33
C: Sulphur

0.90 G: DTP 1.45 1.44


Hardness: 67.489
Results
Elongation at break: 325.062
ZF 11.5 11.5
  Boundaries Elongation at break: 335.107
• Tensile strength-MPA: 11.5-12.0 ZD 325 330
0.60 • Elongation at break-%: 325-335 Hardness 67 67.5
• H-Shore A: 65-67 Hardness: 66.013
Note: Accelerators are preset!


The Design-Expert optimisation graph shows the location of the results 
Boundaries
as a yellow area, but GrafCompounder result is tagged with a flag. • Tensile strength-MPa: 11.5 – 12.0
0.30
5.00 6.00 7.00 8.00 9.00 • Elongation at break-%: 325 – 335
A: ENB • H-Shore A: 65 – 67

6 RFP 3/2013 – Volume 8


is, that a database, which contains such DoE will result in a very good compound [7] Robert O. Babbit, The Vanderbilt Rubber Hand-
unorganised data, cannot be transferred which can serve as a starting point for fur- book, Chapter 9, Compound Design, (1972)
in an organised form needed for a statistic ther evaluation. [8] J. F. Turner, R. J. Kirschbaum, C. P. Louthan,
experimental design software as for instance Profile of Cabot Carbon Blacks in Nitrile Rub-
Design-Expert. A compound prediction with If a compound database is processed by ber, Technical Report RG-125
a DoE software needs a set of data based on the GrafCompounder you will get a com- [9] E. C. McCaffrey, F. C. Church, F. E. Jones, Pro-
planned experiments. pound together with its properties, only lim- file of Cabot Carbon Blacks in Styrene Buta-
ited by the number of the data records con- diene Rubber, Technical Report RG-129
In this paper it could be demonstrated, tained in the database. You will gain knowl- [10] R. J. Kirschbaum, F. E. Jones, Profile of Cabot
that processing the same compound data edge about the influence of ingredients on Carbon Blacks in Natural Rubber, Technical
either based on regression analysis (DoE properties by simulating compounds without Report RG-133
model equations) or based on multiple linear even going into the laboratory, thus saving [11] R. R. Juengel, Profile of Cabot Carbon Blacks
iteration (GrafCompounder type processing) time and money. in Polychloroprene Rubber, Technical Report
leads to results being inside a 95 % confi- RG-134
dence interval and inside the fault tolerances [12] D. C. Novakoski, Profile of Cabot Carbon Blacks
of the testing procedures applicable in the 5. References in EPDM Rubber, Technical Report RG-135
rubber industry. [13] A. C. C. Coolen, A Beginner’s Guide to Mathe-
[1] US Patent 3,781,909 (1973), Edwin Ira Staerns matics of Neural Networks in: Concepts for
The disadvantage of the DoE software [2] European Patent Application 0 647 911 A2 Neural Networks – A Survey, L. J. Landau, J. G.
is its inability to use data being unorga- (1994), Abe Akihito Taylor (Ed.), Springer, Berlin, 1998, 13
nised and outside the scope of an experi- [3] US Patent 6,411,945 B1 (2002), Yukio Nakajima [14] Robert F. Ohm, Rajan G. Vara, Timothy M.
mental design. The advantage of apply- [4] US Patent 6,714,924 (2004), Buckley, Sulfur Cure System Development for
ing a DoE programme is, that one always Craig J. McClahnahan EPDM Produced via Constrained Geometry
gets a very good idea of dependencies [5] US Patent Application 0160114 A1 (2005), Catalyst Technology, Paper presented at the
and interactions between ingredients and Timothy D. Hunt Meeting of the Rubber Division, ACS, Indian-
properties. In addition the output of any [6] Brett Kyle, private information 2001 apolis, IN, May 5 – 8, 1998

RFP 3/2013 – Volume 8 7


HJG Consulting
Rubber & TPE &Coating Technology
Compounding Processing Education

Dr. Hans-Joachim Graf


Josef Leistenschneider Strasse 27
63628 Bad Soden Salmuenster
Germany

Contact:
Phone: +49 (0) 6056 209428 Cell +49 (0) 151 20006394
Telefax: +49 (0) 3222 1520721
www.grafcompounder.com
graf.hans-joachim@t-online.de

View publication stats

You might also like