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

net/publication/257777524

Model-Based Robust Parametric Design of Automatic Cleaning Process

Article  in  Journal of Pharmaceutical Innovation · March 2012

DOI: 10.1007/s12247-012-9120-3

CITATION READS

1 68

3 authors:

Daniela Petrova Elena Koleva

Robotev Bulgarian Academy of Sciences
4 PUBLICATIONS   46 CITATIONS    171 PUBLICATIONS   634 CITATIONS   

SEE PROFILE SEE PROFILE

I. I. Voutchkov
University of Southampton
20 PUBLICATIONS   582 CITATIONS   

SEE PROFILE

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

Optimization of electron beam processing of biopolymer-based materials - Bilateral Cooperation Romanian Academy – Bulgarian Academy of Science View project

automation of an electron beam welding installation View project

All content following this page was uploaded by Daniela Petrova on 06 November 2018.

The user has requested enhancement of the downloaded file.

Model-Based Robust Parametric Design of
Automatic Cleaning Process

Daniela Petrova, Elena Koleva & Ivan

Voutchkov

Journal of Pharmaceutical Innovation

From R&D to Market

ISSN 1872-5120
Volume 7
Number 1

J Pharm Innov (2012) 7:30-37

DOI 10.1007/s12247-012-9120-3

1 23
Your article is protected by copyright and
all rights are held exclusively by Springer
Science+Business Media, LLC. This e-offprint
is for personal use only and shall not be self-
archived in electronic repositories. If you
wish to self-archive your work, please use the
accepted author’s version for posting to your
own website or your institution’s repository.
You may further deposit the accepted author’s
version on a funder’s repository at a funder’s
request, provided it is not made publicly
available until 12 months after publication.

1 23
 Author's personal copy
J Pharm Innov (2012) 7:30–37
DOI 10.1007/s12247-012-9120-3
RESEARCH ARTICLE
Model-Based Robust Parametric Design of Automatic
Cleaning Process
Daniela Petrova & Elena Koleva & Ivan Voutchkov
Published online: 2 February 2012
# Springer Science+Business Media, LLC 2012
Abstract The goal of pharmaceutical industry is to manu-
facture products that meet patients’ needs and expectations,
while satisfying the regulatory requirements. The products
need to meet the required quality and purity characteristics
that are represented to possess. Therefore, in a multi-product
manufacturing facility, appropriately designed cleaning pro-
cesses are essential to avoid cross-contamination between
products and ensure patients’ health and safety. The latest
trend in the development of cleaning validation is using
quality by design methodology (QbD) to determine the most
appropriate parameters of the cleaning processes that will
reduce the risks of cross-contamination. The present study
highlights the model-based approach for robust engineering
design in order to achieve an efficient, reliable, and cost-
effective cleaning process simultaneously.
D. Petrova (*)
Actavis,
3 Samokovsko shose Str,
Dupnitsa 2600, Bulgaria
e-mail: daniella_petrova@abv.bg
D. Petrova
e-mail: dpetrova@actavis.bg
E. Koleva
Institute of Electronics, Bulgarian Academy of Sciences,
72 Tsarigradsko shose,
Sofia 1784, Bulgaria
e-mail: kolevae@ie.bas.bg
I. Voutchkov
Faculty of Engineering and the Environment,
University of Southampton,
Southampton s017 1bj, UK
e-mail: I.I.Voutchkov@soton.ac.uk
Keywords Quality by design . Cleaning development .
Cleaning validation . Robust parametric design . Statistical
modeling . Multiobjective optimization
Introduction
The cleaning of the equipment is one of the most time-
and resource-consuming processes in a multi-product
manufacturing facility. Automatic cleaning methods are
preferred over manual cleaning because they are consid-
ered more reliable and reduce the operator involvement,
which is a generally believed to be a common source of
errors and therefore is difficult to validate. In order to
achieve an efficient cleaning, the input process parame-
ters should be set properly, so to provide the optimal
cleanliness of the production equipment. It is assumed
that an increase in exposure time, temperature, and
detergent concentration will lead to better results, but
it will lead also to excessive costs. A better approach is
to find the optimal parameters where any further in-
crease will not lead to significant change in product
residues on the equipment surfaces after cleaning. In
addition, these settings will obtain the minimal variabil-
ity in their values, which will assure a consistent pro-
cess. This is rather difficult to achieve with a random
adjustment of the parameters; a better practice is to
apply scientifically based methodology such as quality
by design (QbD). Through use of designed experiments
(DoE), modeling, robust design, and optimization procedures
an efficient, consistent, and cost-effective cleaning process
can be developed.
 Author's personal copy
J Pharm Innov (2012) 7:30–37 31

Robust Engineering Design ε is a response noise that incorporates all

other sources of variation;
In industrial processes, input parameters cannot be fixed E denotes mathematical expectation.
precisely but they often exhibit variability and therefore will
vary within some ranges. This can affect the output quality The mean value model of the output characteristic is:
characteristics, causing variations in their levels, which
means that the process would be inconsistent. There are eyðxÞ ¼ E ½yðzÞ ¼ ηðxÞ þ θT EðgÞ
generally two ways to minimize this variance and to im-
The first term ηðxÞ ¼ f T b θ is the output characteristic
prove quality. The first is so called tolerance design, which
model, obtained in laboratory conditions without errors in
consists of tightening the ranges in which the parameters
process parameter levels. The second term takes into ac-
vary, but this is a rather high-cost procedure. The second is
count the variations of the process quality characteristics
called parametric design and is much cheaper and easier to
caused by these errors.
implement. The basic idea is shown in Fig. 1. It shows that  T
there are areas where small variation of input parameters b
θ¼ b θ2 :::b
θ1 b θk is a vector of estimates of the model
cause large variation of the output and others where same coefficients;
variation causes little or no variation of the output charac- f0(f1f2…fk) T
is a vector of the regressors x (known
teristic. We will refer to such area as “robust”. In order to functions of process parameters xi);
improve the quality, the processes/products should be g0h−f is a vector of errors in f;
designed to be insensitive (robust) to these errors by proper h is a vector of the regressors z.
choice of input process parameters values. Another level of
difficulty is introduced when we are searching both for The variance model of the output characteristic is:
optimal and robust solutions, as these are not necessarily  T 
the same. σ2 ¼ σ2 ½yðzÞ ¼ E b θ þ s2" ¼ b
θ yy T b θT Ψ b
θ þ s2" ;
The model-based robust approach for improving the
quality of the process has been successfully applied to where:
different industrial processes [1, 7]. For each of the quality y ¼ g  EðgÞ;
performance characteristics, using polynomial, kriging, or
other mathematical models, two other models are estimated the matrix Ψ0E(ψψT) depends on the structure of the result-
—for their mean values and variances. Quality improvement ing model;
is achieved using some overall criterion or simply by min- s2" is an estimate of the random error of the response.
imization of output characteristic variance, while keeping
their mean values close to the target values.
The measured value of the output characteristic, subject
to errors in input parameters is: Experimental Conditions and Considerations

yðxÞ ¼ E ½yðzÞ þ " Intermediate bulk container (IBC) washing stations are
where: widely used due to their capability to wash and dry all sizes
of IBCs with a minimum of operator involvement, which
x0(x1x2…xm)T is a vector of input parameters; significantly facilitates the cleaning and reduces the opera-
z0x+e is a vector of the real parameters values in tional time. In the initial settings of the washing program
production conditions (with errors e); parameters of our IBC washing station, made by service
engineers, the cleaning was insufficient and the washed
IBCs were still dirty with visible product residues on their
inner and outer surfaces. Furthermore, there were large
variations in data obtained in the Cleaning Validation studies.
These problems led to the need for improvement of the wash-
ing program and the present study highlights the advantages
of applying QbD methodology.
Because access to the IBC washing station and IBCs for
extended experimental studies is quite difficult, initial
experiments were carried out in laboratory using “bench
Fig. 1 Dependence of response’s variation on the operating point scale” model system. Stainless steel coupons AISI 316 L
 Author's personal copy
32 J Pharm Innov (2012) 7:30–37

(10×10 cm), contaminated with some granulation mixture of Table 2 Experimental design and sampling data
a product that is one of the most difficult to clean in practice, Number x1, (%) x2, (°C) x3, (s) y, (μg/100 cm2)
i.e., with very low “Cleanability” were washed using a tablet
disintegration machine DISTEK 3100. An automated clean- 1 1 −1 1 634.49
ing process was simulated in 1,000-ml beakers, while the 2 0 −1 0 1,946.74
input parameters were being controlled. The selected input 3 1 1 −1 16.94
variables are critical parameters that affect the process: 4 0 0 1 7.25
5 1 1 1 10.21
& x1—detergent solution concentration (percent, %);
6 1 −1 −1 1,914.21
& x2—working solution temperature (degrees Celsius, °C);
7 0 0 −1 330.23
& x3—exposure time (seconds, s).
8 −1 −1 −1 7,594.40
Their natural and coded values at the experimental levels 9 −1 1 −1 2,457.73
are listed in Table 1. 10 −1 0 0 968.07
After cleaning, the steel coupons were carefully swabbed 11 −1 −1 1 4,449.67
with appropriate solvent and the samples were analyzed by a 12 −1 1 1 1,312.12
validated HPLC method. For this study, the measured output 13 1 0 0 26.61
quality characteristic was the active substance residues 14 0 1 0 5.85
(micrograms per 100 cm2) that remained on the coupon sur-
face after cleaning.
In order to study the dependence of the active substance Polynomial Regression
residue levels on the input parameters, a three-level D-optimal
experimental design for a second-order polynomial was Regression analysis is a statistical method using polyno-
implemented. These plans allow selecting the samples in such mials to build mathematical models that define quantitative
a way that the determinant of information matrix XTX is relationship between the output characteristic and input
maximized. It is related to the accuracy of the least square parameters [1]:
estimates of the regression coefficients, which is defined as:
   by ¼ b
θT f
1 2
Var b θ ¼ XTX σ
The estimations of the unknown coefficients of the poly-
where σ is the variance of the estimate error. It can be seen
2 nomial model can be obtained by the least square method:
that to improve the quality of the fit, the determinant of  1
b
θ ¼ FT F FT y
information matrix XTX should be maximized. Thus, D-
optimality provides a low variance estimates and ensures where:0 f11    f1k 1
maximum accuracy of the estimated regression coefficients. B . .. . C
F ¼ @ .. . .. A is the extended design matrix,
The randomized experimental design in coded form and
the sampling data are presented in Table 2. fn1    fnk called matrix of regressors;
y0(y1y2…yn)T is the response vector.

Modeling of Active Substance Residue Levels The use of the classical linear regression analysis is
After Cleaning limited by the assumptions of the normally distributed,
independent, and homogeneous in variance output random
Two different modeling methods were used to fit the output noise, and the requirement of the experiments implementa-
characteristic: polynomial regression and kriging. tion without any errors, which is not always possible.
At implementation of the regression analysis, a better
Table 1 Natural and model was obtained for the response after undertaking a
coded process Natural Coded non-linear transformation (ln(y)) of the data values. This
parameters simplifies and improves the model:
−1 0 1

x1 1.0 3.0 5.0 cyðxÞ ¼ 3:602  1:675x1  2:904x2  1:909x3

ln
x2 20.0 45.0 70.0
0:813x1 x2 þ 1:477x21 þ 1:069x22
x3 60 90 120 þ0:289x23 þ 1:503x21 x2 þ 1:563x21 x3
 Author's personal copy
J Pharm Innov (2012) 7:30–37 33

All of the estimated regression coefficients are significant parameters θj and pj by the next distance measure:
and the polynomial model is adequate with determination !
h    i X
k  p
coeficients as follows:  ðiÞ ðlÞ  j
cor Y xðiÞ ; Y xðlÞ ¼ exp  θj xj  xj  þl
j¼1
R ¼
2
0:99825; R2adj ¼ 0:99433 and R2pred ¼ 0:98909

The initial levels of contamination were not exactly where:

equal during the different runs. They depended also on θj is a measure of the “activity” of the approximated function;
the particular formulation, dirty equipment hold time pj is related to the “smoothness” of the correlation
(DEHT), etc. This will introduce some noise, which will depending on the distance;
affect the response. Therefore, the use of classical linear λ is a regularization term, added to filter the noise and
regression analysis is not quite correct, since the as- prevent the overfitting of the sampled data.
sumption for constant variance of the output random
error is violated and the obtained polynomial model The correlation matrix of all the observed data is:
might be biased. A better model can be obtained by
0   ð1Þ   ð1Þ       1
using a suitable approximation method, which takes into cor y x ; y x . . . cor y xð1Þ ; y xðnÞ
account that noise. B .. .. .. C
Ψ ¼@ . . . A
  ðnÞ   ð1Þ    ðnÞ   ðnÞ 
cor y x ; y x . . . cor y x ; y x
Kriging
The kriging parameters can be obtained by maximum
Kriging [2–5] is an interpolative approximation method likelihood estimation. The likelihood function of y:
for modeling that makes predictions based on an expo- " #
nentially weighted sum of a known sample of data. It is 1 ðy  1μÞT Ψ 1 ðy  1μÞ
LðyÞ ¼ n 1 exp 
a form of radial basis functions which has more tuning ð2pσ2 Þ 2 jΨ j 2 2σ2
parameters and hence is suitable for problems with
higher complexity and lower number of experiments. is taken a natural logarithm to simplify and the estimates of
Kriging models are very flexible; depending on its μ and σ2 are obtained by taking derivatives of ln(L(y)):
1T Ψ 1 y
parameters they can either provide an exact interpolation b¼
μ 1T Ψ 1 1
T 1
of the data, or smooth the data in the presence of b2 ¼ ðy1μÞ Ψn ðy1μÞ
σ
numerical noise. Kriging may work well in noisy con- These estimates are substituted back in ln(L(y)), the con-
ditions under the assumption that the noise and under- stant members are removed and resulted concentrated ln-
lying signal are additive. It also provides a measure of likelihood function
the error or uncertainty at the unsampled locations
which could be used to determine the location of further n  2 1
CLF ¼  ln σ
b  lnjΨ j
experiments. Kriging works and with correlated data, 2 2
incorporates the spatial correlation of the data. It can is maximized by searching the function using a global
work also when observations are independent and then search method (e.g. a genetic algorithm) to obtain the esti-
its estimation coincides with that determined by using mates of the parameters θj, pj, and λ.
least-squares regression. To estimate an unobserved response at a new point x, the
Implementing the experimental design at a set of input experimental data y are augmented with the new prediction
 T T
T
parameters X ¼ xð1Þ ; xð2Þ ; . . . ; xðnÞ , the observed responses y
0  ð1Þ  1 by : ey ¼ and a vector of correlations between the
y x by
B . C
y ¼ @ .. A can be expressed as: new prediction and the previously sampled data is formed:
  0   ð1Þ  1
y xðnÞ cor y x ; yðxÞ
B .. C
y ¼ μ þ "ðxÞ y¼@ . A
  ðnÞ 
where μ is the mean of the output responses and ε(x) is a cor y x ; yðxÞ
Gaussian random function with zero mean and variance σ2. It
is not assumed that the ε is constant as in the regression If the unobserved point x is close to x(i), then these points
analysis, but these errors are correlated. The correlation be- are correlated and the predicted response will be strongly
tween two responses at points x(i) and x(l) depends on the influenced by the response at x(i), whereas the points move
absolute distance between these points and the correlation apart, the correlation is small and the predicted response will
 Author's personal copy
34 J Pharm Innov (2012) 7:30–37

only be weakly influenced by the response at x(i). The Table 4 The tolerance intervals of the input parameters
augmented correlation matrix is: Process parameters Tolerance ranges Coded variances

e Ψ y
Ψ¼ Concentration, (%) x1 ±0.3 0.0025
yT 1
Temperature, (°C) x2 ±5 0.0044
The prediction at this new point x is obtained by maxi- Contact time, (s) x3 ±5 0.0031
mizing the likelihood function of the augmented data
" #
1 e 1 ðey  1b
μÞT Ψ
ðey  1b μÞ and kriging models, were estimated. Contour plots for both
LðeyÞ ¼ exp 
  n   12 2bσ2 statistical models as a function of two input parameters
σ2 2 Ψ
2pb e
(detergent concentration and temperature) are presented in
by differentiating with respect to by and setting the derivative Figs. 2 and 3 (for polynomial regression) and Figs. 4 and 5
to zero: (for kriging). The third parameter (exposure time) is a con-
stant at its basic level (90 s). The values of the active
b þ y T Ψ 1 ðy  1b
byðxÞ ¼ μ μÞ substance residues per 100 cm2 and the variances in the
contour plots are in their logarithmic form (ln).
The estimates of kriging parameters, obtained using sta-
The multiobjective optimization will provide robustness of
tistical software QstatLab (http://www.qstatlab.co.uk) are
the cleaning process (less sensitivity towards the errors in
presented in Table 3.
input parameters) and will assure the desired level of cleanli-
ness simultaneously. It is concerned with the minimization of
a vector of objectives [6]:
Robust Parametric Design of the IBCs Washing Process
min FðxÞ ¼ min½f1 ðxÞ; f2 ðxÞ
x2Γ x x2Γ x
During the washing cycle in IBC washing station, the input
where:
parameters: detergent concentration, temperature of the so-
lution, and contact time vary in some tolerance ranges, f1(x) is a function of the mean value of the active
expressed as: substance residues after cleaning;
f2(x) is a function of the variance of active substance
ð xi  ς i ; xi þ ς i Þ
residues.
The tolerance ranges of the parameters, listed in func- When the extrema of the two functions do not coincide,
tional requirement specification of the IBC washing station, because of the competition of the two components of F(x),
are given in Table 4. To be used in mean and variance model
building, they were converted into coded form by the next
formula, based on the assumption of normal distribution of
the errors in the input parameters:
ς 2i
σ2i ¼
9w2i
where:
ςi is the half interval of parameter’s tolerance range
ωi is the half interval of parameter variation during the
experiments.

Using the robust engineering design methodology, the

models of the mean value and the variance of the active
substance residues, based on both polynomial regression

Table 3 Kriging model

parameters Number θj pj λ

1 10−0.2495 2 10−3.8125
2 10−0.6176 2
3 10−1.0476 2 Fig. 2 Contour plot of the mean value function based on polynomial
regression model
 Author's personal copy
J Pharm Innov (2012) 7:30–37 35

Fig. 3 Contour plot of the variance function based on polynomial Fig. 5 Contour plot of the variance function based on kriging model
regression model

point of the output characteristic’s mean value function

there is no unique solution but a set of noninferior solutions,
coincides with the stationary point of the variance function
called Pareto front (Fig. 6), in which the improvement in one
and it always corresponds to the minimal variance [1]. It is
objective causes a degradation of another.
confirmed in the case of kriging, which filters better the
In the contour plots (Figs. 2, 3, 4, and 5) can be seen that
noise, arising from different levels of initial contamination,
the minima of the both functions (the mean value and the
while in case of regression models there is a little shift. It
variance of the active substance residues) are very close to
can be seen also in the listed optimal solutions (Table 5),
each other for polynomial regression models, as in the case
which are a part of Pareto front, obtained by implemented
of kriging models they coincide. It can be proved, that in
multiobjective optimization using genetic algorithm. Using
presence of errors in input parameters only, the stationary
kriging models, the optimal input parameters and the
resulted output characteristics are almost the same, while
using regression models, the results are different (Table 5,
Figs. 7 and 8).

Fig. 4 Contour plot of the mean value function based on kriging Fig. 6 Pareto front, illustrating the trade-off between the two
model objectives
 Author's personal copy
36 J Pharm Innov (2012) 7:30–37

Variance
Table 5 Pareto optimal solutions 0.120

Number x1 x2 x3 by b2
σ

Regression model-based optimization results 0.100

1 3.55 70.0 120 0.831 1.0565

2 3.95 69.7 119 1.035 1.0541
3 3.89 69.8 120 0.967 1.0541 0.080

4 3.77 70.0 119 0.902 1.0549

Kriging model-based optimization results
0.060
5 3.68 63.2 116 1.474 1.0006
6 3.66 62.8 116 1.476 1.0006
7 3.66 63.1 116 1.475 1.0006
0.040
8 3.67 62.9 116 1.475 1.0006

Models Testing and Validation 0.020

Validation experiments at two optimal operating points (1

0.000
and 5 from Table 5) were carried out at industrial conditions.
The results are given in Table 6. 2.0 4.0 6.0 8.0

The 95% confidence intervals for mathematical expecta- Mean

tion of validating experiments are calculated by: Fig. 8 Pareto front, obtained using multiobjective optimization, based
on kriging models (values are in ln)
σc
b σ
bc
bc  tT pffiffiffiffi  μ  μ
μ bc þ tT pffiffiffiffi
N N
σ
bc is the estimates of standard deviation;
where: N is the number of validating experiments;
tT(α,ν) is a critical value of t distribution determined by
μ is the mathematical expectation of the response;
α00.05 significance level and ν0N−1 degrees of
bc
μ is the estimates of mathematical expectation (mean
freedom.
value);
Variance The obtained 95% confidence intervals are as follows:
& for validating experiments, carried out at the regression
0.200
optimal conditions: 1.97–5.11;
& for validating experiments, carried out at the kriging
optimal conditions: 0.58–4.11.
Kriging predicted optimal values fall within the calculat-
0.150 ed confidence interval, while the regression ones do not.
A root mean squared error (RMSE) was used for models
testing:
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
uP  2
u n ðiÞ
t i¼0 yðiÞ  by
0.100 RMSE ¼
N

Table 6 Validation
experiments y1 y2 y3 y4
0.050
At regression optimal operating conditions
0.0 2.0 4.0 6.0 8.0
Mean 4.09 2.39 3.10 4.59
At kriging optimal operating conditions
Fig. 7 Pareto front, obtained using multiobjective optimization, based 3.87 2.44 1.41 1.65
on regression models (values are in ln)
 Author's personal copy
J Pharm Innov (2012) 7:30–37
& for model based on the regression RMSE02.84;
& for model based on kriging RMSE01.29.
From model testing and calculated confidence intervals,
it is obvious that kriging is the method that is more accurate
in the conditions under the experiments carried out.
Conclusion
The study presented a QbD approach for quality improve-
ment of products in pharmaceutical manufacturing via ap-
propriate designed cleaning processes.
Given the specificity of the cleaning processes, the use of
classical linear regression analysis is not always the best
choice, since some of its assumptions can be violated. The
experiments are often carried out in presence of noise
(DEHT; the particular formulation; the number of batches,
produced on the equipment, etc. are different), which results
in unequal initial levels of equipment surface contamination
and it in turn affect the final results and models. For the
considered cleaning process, kriging is the more reliable
method for modeling and robust optimization, as its predic-
tion appears to be more accurate than the standard regres-
sion model.
The dependence of the product residues, which remain on
the equipment surfaces after cleaning, on the input process
parameters is not a pure linear function, but it has an optimum,
where any further increase in parameters will not result in
significant change in the output product residues on the equip-
ment surfaces. Using robust parametric design, the chosen
optimal conditions will also assure the minimal variance in
their values. Therefore it is possible to design a reliable,
robust, and cost-effective cleaning method simultaneously.
For this particular cleaning validation study, based on the
validation data at kriging optimal operating conditions, we
View publication stats
37
achieved a 74.4% reduction of the average value of the
active substance residues; a 99.4% reduction of the variance,
and a 28.9% cost reduction. In addition, the development
time was reduced from more than 3 months to less than
2 weeks and the reduction in validation time was 90%.
The QbD approach provides the following benefits:
& reduced development time and costs;
& efficient, effective, and consistent cleaning methods;
& cost-effective cleaning processes;
& high cleaning assurance level;
& quickly achieved validated status of cleaning processes.
Acknowledgments The authors gratefully acknowledge M. Paleva
for her assistance in analytical testing and Prof. I. Vuchkov for his
valuable comments on this work.
References
1. Vuchkov I, Boyadjieva L. Quality improvement with design of
experiment. The Netherlands: A. Keller, Kluwer Academic Publish-
ers; 2001.
2. Forrester A, Sobester A, Keane A. Engineering design via surrogate
modelling. New York: Wiley; 2008.
3. Li J, Heap A. A review of spatial interpolation methods for envi-
ronmental scientists. Canberra: Geoscience Australia; 2008. Record
2008/23.
4. Jiam Z, Davis E, Muzzio F, Ierapetritou M. Predictive modeling for
pharmaceutical processes using kriging and response surface. J
Pharm Innov. 2009;4:174–86.
5. Smith R. Kriging with estimated parameters. UNC Statistics and
Operations Research Colloquium January 10 2005.
6. Steuer R. Multiple criteria optimization: theory, computations, and
application. New York: Wiley; 1986.
7. Koleva E, Mladenov G, Vuchkov I, Velev K, Lamond B, Petrova D.
Program tool EBMR-MREDO for robust engineering quality opti-
mization of EBMR of titanium, E+E, 5–6; 2009, 252–259.