Composite Structures 93 (2011) 1309–1310

Contents lists available at ScienceDirect

Composite Structures
journal homepage: www.elsevier.com/locate/compstruct

Note

Comment on ‘‘Hybrid computational strategy based on ANN and GAPS:

Application for identification of a non-linear model of composite material’’
W. Sha ⇑
School of Planning, Architecture and Civil Engineering, Queen’s University Belfast, Belfast BT7 1NN, UK

a r t i c l e i n f o a b s t r a c t

Article history: This short communication comments on a series of papers using artificial neural networks published by
Available online 13 November 2010 Guessasma and co-workers in structures journals. The issues discussed include the size of the database
for training a neural network, database enlargement for training a neural network, and extrapolation
Keywords: based on modelling results. The paper includes criticisms of other mistakes in those previously published
Neural networks papers.
Non-linear model Ó 2010 Elsevier Ltd. All rights reserved.
Size of database
Database enlargement
Extrapolation

1. Introduction 2. Size of the database and database enlargement for training a

neural network
In 2009 and 2006, Guessasma and co-workers published a
series of papers in structures journals, using artificial neural net- As shown in [2, Table 1], there is very limited variation of data.
work modelling [1–3]. The way that neural network modelling is There are just three values for each input parameter in the entire
used is similar in all these papers, although there are two differ- experimental data set. The use of neural network (NN) is not
ent modelled phenomena. Papers [2,3] are very similar, near appropriate for such small database, as three values for a variable
duplication in many places, bordering duplicate publication. are not able to reveal relationships other than the simplest ones.
Although quoting the most recent of them in the title, this This was further explained in a separate paper [4].
comment paper attempts to comment on these three papers When training NN with databases simple in nature and limited
collectively. in size such as that used in [2,3], the NN will give straightforward
There are three major issues in the papers by Guessasma and answers, confirming the unnecessity of using it in the first place.
co-workers as far as modelling using neural networks is concerned. This is shown in [3, Figs. 4b and 5b]. If the original experimental
The first is about the need and effectiveness of developing neural data jumps from stable to unstable (discreet), but a neural network
network models based on very few experimental data points, the of a continuous function is used, the NN is bound to give something
second is about the database enlargement for training neural net- similar to linear regression, as shown in these figures.
works, and the third is on the extrapolation based on neural net- When carrying out database enlargement, an uncertainty of ±3%
work models developed. These three problems are related, in for each parameter was used in [2]. No explanation is given as to
that the second and the third problems would not arise if there why this error was assumed, but it is important for creating extra
were no first problem, i.e., if there were sufficiently large databases data. In practice, it is unlikely for different parameters to have ex-
to train the neural networks. actly the same relative error percentage. Similarly, in [3], the
In the following text, when quoting figures, table and equations uncertainty of 2.5% for each parameter was used, without explana-
in papers by Guessasma and co-workers, they are quoted within tion on how it was determined and why a value different from [2]
square brackets, after reference numbers, in order to show that was adopted for the same input variables.
they do not belong to this comment paper.
3. Extrapolation based on modelling results

It is stated in [2] that ‘‘One of the advantages of using the opti-

⇑ Tel.: +44 28 90974017; fax: +44 28 90974278. mised neural network structure is to enlarge the parameter uncer-
E-mail address: w.sha@qub.ac.uk tainties considered in the experiments and the finite element

0263-8223/$ - see front matter Ó 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compstruct.2010.11.003
1310 W. Sha / Composite Structures 93 (2011) 1309–1310

analysis (3%) and predict the performance for this larger domain’’,  In [1, Fig. 7], comparison between NN calculations with training
implying extrapolation. However, the later discussion about [2, data is made. Such comparison is not useful because NN is
Fig. 5] shows the lack of extrapolation ability, as expected, voting known to overfit the training data when the database is small.
against the authors’ own statement of ‘‘advantages’’. Extrapolation Comparison should be made between NN calculations with
should not be used. Controversially again, the authors conclude the so-called testing data, i.e., data not used during the training
that ‘‘This result proved the extrapolation capability of the ANN process.
as it permitted us to predict instability situations outside the inter-  In [3, Figs. 2 and 4–6], the two sub-figures of a and b in each fig-
vals adopted for the databases’’, against their own findings. The ori- ure are placed the wrong way round.
ginal 3% is extended to 5–10%, but no explanation is given on how  The final comment is on the number of epochs selected by
these larger percentages are obtained. There is no validation of Guessasma and co-workers (different in each of their papers)
accuracy or even estimation of it. Later in that paper, it is stated, that is chosen to ‘‘avoid overtraining’’ of the ANN. In [1],
‘‘When using database 1 for the optimisation, the decrease of per- 10,000 epochs are used and it would have been important to
formance (increase of quadratic function) is overestimated. The show that this number is indeed low enough to prevent
scatter between the two curves relative to the databases used for overtraining.
the optimisation of the neural network was on average 4.38%,
which is reasonable.’’ Again, there is no explanation or justification
on why it is overestimated and why 4.38% is reasonable. References

[1] Bassir DH, Guessasma S, Boubakar L. Hybrid computational strategy based on

4. Other comments ANN and GAPS: Application for identification of a non-linear model of
composite material. Compos Struct 2009;88:262–70.
[2] Fall H, Guessasma S, Charon W. Prediction of stability and performance of an
This section lists comments not related to the main problems active mechanical structure under uncertainty conditions using finite element
detailed above, but examples of bad paper writing and research and neural computation. Eng Struct 2006;28:1787–94.
[3] Fall H, Guessasma S, Charon W. Stability analysis of a controlled aluminium
practice. panel using neural network methodology. Comput Struct 2006;84:835–42.
[4] Sha W. Comment on ‘‘Modeling of tribological properties of alumina fiber
 In [1], it is stated ‘‘ANN is known to provide an implicit form of reinforced zinc–aluminum composites using artificial neural network’’ by K.
Genel et al. [Mater. Sci. Eng. A 363 (2003) 203]. Mater Sci Eng A
the correlations between inputs and outputs of a given problem 2004;372:334–5.
whatever is their complexity and parameter interdependency’’. [5] Kutuk MA, Atmaca N, Guzelbey IH. Explicit formulation of SIF using neural
The description of ‘‘implicit’’ is incorrect. An artificial neural networks for opening mode of fracture. Eng Struct 2007;29:2080–6.
[6] Okuyucu H, Kurt A, Arcaklioglu E. Artificial neural network application to the
network (ANN) model gives explicit relations between input friction stir welding of aluminium plates. Mater Des 2007;28:78–84.
variables and output functions, in the form of [1, Eqs. (1) and
(2)]. By default, explicit relations (equations) are given by a Wei Sha obtained a BEng at Tsinghua University in 1986. He was awarded in 1992 a
trained neural network between the inputs and the outputs PhD by Oxford University and in 2009 a DSc by Queen’s University Belfast. He has
for the system considered. These equations are normally not worked at Imperial College, Cambridge University and Queen’s University Belfast.
He is presently Professor of Materials Science, with research interests in titanium
shown, because they usually do not have any physical explana-
alloys, mathematical models, phase transitions, microstructure, reaction kinetics,
tions and they are rather long, but they do exist. In fact, such computer simulation, neural networks, and nitriding. He is a Fellow of The Institute
equations are occasionally included in NN papers, for examples, of Materials, Minerals & Mining (FIMMM) and a Fellow of The Institute of Metal
by Kutuk et al. [5] and Okuyucu et al. [6]. Finishing (FIMF).